Secure processors must defend against a wide array of microarchitecture side-channels, including those induced by a shared memory controller. Multiple studies have proposed techniques that allocate ''turns'' (within the memory controller) to each co-scheduled virtual machine (VM), and introduce gaps between VM turns to prevent resource conflicts and side-channels. In spite of past advancements in secure memory scheduling, the elimination of side-channels imposes a performance slowdown of 2x. We observe that one of the causes of this slowdown is that the memory controller schedule accommodates the worst case, i.e., it is prepared to handle either reads or writes. The key insight in this work is that the schedule can be more efficient if we designate every turn to handle fixed patterns of reads and writes.In particular, we introduce a read-optimized turn and a write-optimized turn. Coarse-grain application profiling helps determine how often the two types of turns are invoked, without leaking sensitive information. We also add flexibility so that a read-optimized turn can opportunistically also issue writes, and vice versa. This provides a good balance between restrictions and flexibility; between throughput and utilization. The proposed FlexProf memory controller improves performance by up to 33% with a geometric mean gain of 8% on mixed workloads, relative to state-of-the-art methods. Over half the memory-intensive programs evaluated exhibit performance gains of over 10%.

Data prefetching is vital in high-performance processors and a large body of research has introduced a number of different approaches for accurate prefetching: stride detection, address correlating prefetchers, delta pattern detection, irregular pattern detection, etc. Most recently, a few works have leveraged advances in machine learning and deep neural networks to design prefetchers. These neural-inspired prefetchers observe data access patterns and develop a trained model that can then make accurate predictions for future accesses. A significant impediment to the success of these prefetchers is their high implementation cost, for both inference and training. These models cannot be trained in real-time, i.e., they have to be trained beforehand with a large benchmark suite. This results in a large model (that increases the overhead for inference), and the model can only successfully predict patterns that are similar to patterns in the training set.

Deep neural networks are widely used in a range of commercial services. Many of these services are hosted on the cloud, requiring users to send their personal data to the cloud. This, in turn, exposes the user’s private and sensitive data to several third parties. To address this problem, Homomorphic Encryption (HE) has been introduced, where the user encrypts their data before sending it to the cloud; the cloud performs operations on encrypted data and returns a ciphertext that the user must then decrypt. While this approach keeps user data private, it demands orders of magnitude more computation and data movement. It is, therefore, imperative to design hardware/software techniques to lower the overheads when executing AI services under Homomorphic Encryption schemes.In this paper, we consider a range of HE implementations for AI inference and address the key bottlenecks in state-of-the-art frameworks. We start by making the case for a hybrid HE and Multi-Party Computation (MPC) scheme that is more practical than pure Fully HE. This paper introduces new techniques at various levels: (i) we introduce new data packing techniques that result in lower data movement, (ii) we introduce new dataflows that increase reuse and reduce other costly HE operations (rotations, key switching, NTT conversion), (iii) we evaluate Hyena on a balanced pipelined architecture that efficiently handles the above primitives. The resulting framework, Hyena (new packing + dataflow), achieves better performance and energy than several packing baselines. Compared to the widely used Channel-packing, Hyena is 38× faster and achieves 162× lower energy consumption, with an overall ResNet20 inference end-to-end latency of 11.4 ms, using a 163 mm2accelerator dissipating 16.75 W.

Several deep neural network (DNN) accelerators have been designed to exploit the sparsity exhibited by DNN activations and weights. State-of-the-art sparse accelerators can be described as either Pixel-first or Channel-first accelerators, each with its unique dataflow and compression format aiding its dataflow. The former expends significant energy updating neuron partial sums, while the latter expends significant energy in handling the index metadata. This work introduces a novel microarchitecture and dataflow that reconciles these trade-offs by adopting a Pixel-first compression and Channel-first dataflow. The proposed microarchitecture has a simpler index-generation logic combined with an accumulator buffer hierarchy and crossbar with low wiring overhead. The compression format and dataflow promote high temporal locality in neuron updates, further lowering energy. Finally, we introduce work partitions across processing elements that naturally lead to load balance without offline analysis. Compared to four state-of-the-art baselines, the proposed architecture, CANDLES, significantly outperforms three and matches the performance of the fourth. In terms of energy, CANDLES is between 2.5× and 5.6× more energy-efficient than these four baselines.

Oblivious RAMs (ORAMs) are proposed to completely hide access patterns. However, most ORAM constructions are expensive and not suitable to deploy in a database for supporting query processing over large data. In this work, we design a practical oblivious query processing framework to enable efficient query processing over a cloud database. In particular, we focus on processing multiple range and kNN queries asynchronously and concurrently with high throughput. The key idea is to integrate indices into ORAM which leverages a suite of optimization techniques (e.g., oblivious batch processing and caching). Our construction shows an order of magnitude speedup in comparison with other baselines over large datasets.

Increasingly, individuals and companies adopt a cloud service provider as a primary data and IT infrastructure platform. The remote access of the data inevitably brings the issue of trust. Data encryption is necessary to keep sensitive information secure and private on the cloud. Yet adversaries can still learn valuable information regarding encrypted data by observing data access patterns. To solve such problem, Oblivious RAMs (ORAMs) are proposed to completely hide access patterns. However, most ORAM constructions are expensive and not suitable to deploy in a database for supporting query processing over large data. Furthermore, an ORAM processes queriessynchronously, hence, does not provide high throughput forconcurrent query processing. In this article, we design a practicaloblivious query processing frameworkto enable efficient query processing over a cloud database. In particular, we focus on processing multiple range and$k$NN queriesasynchronously and concurrently with high throughput. The key idea is to integrate indices into ORAM which leverages a suite of optimization techniques (e.g., oblivious batch processing and caching). The effectiveness and efficiency of our oblivious query processing framework is demonstrated through extensive evaluations over large datasets. Our construction shows an order of magnitude speedup in comparison with other baselines.

Oxford Nanopore Technologies (ONT) MinION is a pocket-sized portable DNA sequencer for on-the-field DNA sequencing. The ONT data analysis pipeline requires substantial amounts of computational resources for the basecalling and read alignment step in the pipeline. We study the compute and memory requirement of the basecalling step and present the detailed microarchitecture for accelerating it on FPGA. We also make the case that an embedded FPGA-based SoC with partial reconfiguration (PR) capability is an ideal substrate for accelerating ONT data analysis pipeline, due to the diversity in these application kernels and the evolving nature of the algorithms. Focusing on two widely used ONT data analysis pipelines, we present two modes of PR operation to cater to either latency or throughput requirements of these two pipelines and demonstrate how to amortize the partial reconfiguration cost, overlap the kernels, minimize the data transfer cost and achieve 111× higher throughput and 6× reduction in latency using these two PR modes.

As technologies continue to shrink, memory system failure rates have increased, demanding support for stronger forms of reliability. In this work, we take inspiration from the two-tier approach that decouples correction from detection and explore a novel extrapolation. We propose Dvé, a hardware-driven replication mechanism where data blocks are replicated in 2 different sockets across a cache-coherent NUMA system. Each data block is also accompanied by a code with strong error detection capabilities so that when an error is detected, correction is performed using the replica. Such an organization has the advantage of offering two independent points of access to data which enables: (a) strong error correction that can recover from a range of faults affecting any of the components in the memory, upto and including the memory controller, and (b) higher performance by providing another nearer point of memory access. Dvé realizes both of these benefits via Coherent Replication, a technique that builds on top of existing cache coherence protocols for not only keeping the replicas in sync for reliability, but also to provide coherent access to the replicas during fault-free operation for performance. Dvé can flexibly provide these benefits on-demand by simply using the provisioned memory capacity which, as reported in recent studies, is often underutilized in today’s systems. Thus, Dvé introduces a unique design point that offers higher reliability and performance for workloads that do not require the entire memory capacity.

Modern workloads such as neural networks, genomic analysis, and data analytics exhibit significant data-intensive phases (low compute to byte ratio) and, as such, stand to gain considerably by using processing-in-memory (PIM) solutions along with more traditional accelerators. While PIM has been researched extensively, the granularity of computation offload to PIM and the granularity of memory access arbitration between host and PIM, as well as their implications, have received relatively little attention. In this work, we first introduce a taxonomy to study the design space whilst considering these two aspects. Based on this taxonomy, we observe that much of PIM research to date has largely relied on coarse-grained approaches which, we argue, have steep costs (incompatibility with mainstream memory interfaces, prohibition of concurrent host accesses, and more). To this end, we believe that better support for fine-grained approaches is warranted in accelerators coupled with PIM-enabled memories.

Spiking neural networks (SNNs) are expected to be part of the future AI portfolio, with heavy investment from industry and government, e.g., IBM TrueNorth, Intel Loihi. While Artificial Neural Network (ANN) architectures have taken large strides, few works have targeted SNN hardware efficiency. Our analysis of SNN baselines shows that at modest spike rates, SNN implementations exhibit significantly lower efficiency than accelerators for ANNs. This is primarily because SNN dataflows must consider neuron potentials for several ticks, introducing a new data structure and a new dimension to the reuse pattern. We introduce a novel SNN architecture, SpinalFlow, that processes a compressed, time-stamped, sorted sequence of input spikes. It adopts an ordering of computations such that the outputs of a network layer are also compressed, time-stamped, and sorted. All relevant computations for a neuron are performed in consecutive steps to eliminate neuron potential storage overheads. Thus, with better data reuse, we advance the energy efficiency of SNN accelerators by an order of magnitude. Even though the temporal aspect in SNNs prevents the exploitation of some reuse patterns that are more easily exploited in ANNs, at 4-bit input resolution and 90% input sparsity, SpinalFlow reduces average energy by 1.8×, compared to a 4-bit Eyeriss baseline. These improvements are seen for a range of networks and sparsity/resolution levels; SpinalFlow consumes 5× less energy and 5.4× less time than an 8-bit version of Eyeriss. We thus show that, depending on the level of observed sparsity, SNN architectures can be competitive with ANN architectures in terms of latency and energy for inference, thus lowering the barrier for practical deployment in scenarios demanding real-time learning.

The memory system is vulnerable to a number of security breaches, e.g., an attacker can interfere with program execution by disrupting values stored in memory. Modern Intel® Software Guard Extension (SGX) systems already support integrity trees to detect such malicious behavior. However, in spite of recent innovations, the bandwidth overhead of integrity+replay protection is non-trivial; state-of-the-art solutions like Synergy introduce average slowdowns of 2.3× for memory-intensive benchmarks. Prior work also implements a tree that is shared by multiple applications, thus introducing a potential side channel. In this work, we build on the Synergy and SGX baselines, and introduce three new techniques. First, we isolate each application by implementing a separate integrity tree and metadata cache for each application; this improves metadata cache efficiency and improves performance by 39%, while eliminating the potential side channel. Second, we reduce the footprint of the metadata. Synergy uses a combination of integrity and error correction metadata to provide low-overhead support for both. We share error correction metadata across multiple blocks, thus lowering its footprint (by 16×) while preventing error correction only in rare corner cases. However, we discover that shared error correction metadata, even with caching, does not improve performance. Third, we observe that thanks to its lower footprint, the error correction metadata can be embedded into the integrity tree. This reduces the metadata blocks that must be accessed to support both integrity verification and chipkill reliability. The proposed Isolated Tree with Embedded Shared Parity (ITESP) yields an overall performance improvement of 64%, relative to baseline Synergy.

Applications in the cloud are vulnerable to several attack scenarios. In one possibility, an untrusted cloud operator can examine addresses on the memory bus and use this information leak to violate privacy guarantees, even if data is encrypted. The Oblivious RAM (ORAM) construct was introduced to eliminate such information leak and these frameworks have seen many innovations in recent years. In spite of these innovations, the overhead associated with ORAM is very significant. This paper takes a step forward in reducing ORAM memory bandwidth overheads. We make the case that, similar to a cache hierarchy, a lightweight ORAM that fronts the full-fledged ORAM provides a boost in efficiency. The lightweight ORAM has a smaller capacity and smaller depth, and it can relax some of the many constraints imposed on the full-fledged ORAM. This yields a 2-level hierarchy with a relaxed ORAM and a full ORAM. The relaxed ORAM adopts design parameters that are optimized for efficiency and not capacity. We introduce a novel metadata management technique to further reduce the bandwidth for relaxed ORAM access. Relaxed ORAM accesses preserve the indistinguishability property and are equipped with an integrity verification system. Finally, to eliminate information leakage through LLC and relaxed ORAM hit rates, we introduce a deterministic memory scheduling policy. On a suite of memory-intensive applications, we show that the best Relaxed Hierarchical ORAM (ρ) model yields a performance improvement of 50%, relative to a Freecursive ORAM baseline.

Real-time phase detection enables dynamic adaptation of systems based on different program behavior. Many phase detection techniques have been proposed, with the most successful relating the phases back to application code. In the scope of online phase detection, techniques employ sampling to mitigate the overheads of the phase detection framework. When phase intervals are long enough, sampling approaches perform well. We reopen the question of phase interval length by performing in-depth analysis on the trade-offs between overhead and phase detector performance. We present a new metric which captures the statistical trade-off between phase interval length, phase stability, and the number of phases. We find that while shorter phases perform best in the context of online optimization, existing implementations suffer from performance degradation and overhead at shorter interval sizes. To address this gap, we present the Precise Online Phase (POP) detector. The POP detector utilizes performance counters to build signatures, which are virtually lossless at finer granularity. As a second order, the simplicity of the detector reduces the runtime overhead to just 1.35% and 0.09% at 10M and 100M instruction intervals, respectively.

In spite of several recent advancements, data movement in modern CNN accelerators remains a significant bottleneck. Architectures like Eyeriss implement large scratchpads within individual processing elements, while architectures like TPU v1 implement large systolic arrays and large monolithic caches. Several data movements in these prior works are therefore across long wires, and account for much of the energy consumption. In this work, we design a new wire-aware CNN accelerator, WAX, that employs a deep and distributed memory hierarchy, thus enabling data movement over short wires in the common case. An array of computational units, each with a small set of registers, is placed adjacent to a subarray of a large cache to form a single tile. Shift operations among these registers allow for high reuse with little wire traversal overhead. This approach optimizes the common case, where register fetches and access to a few-kilobyte buffer can be performed at very low cost. Operations beyond the tile require traversal over the cache's H-tree interconnect, but represent the uncommon case. For high reuse of operands, we introduce a family of new data mappings and dataflows. The best dataflow, WAXFlow-3, achieves a 2× improvement in performance and a 2.6-4.4× reduction in energy, relative to Eyeriss. As more WAX tiles are added, performance scales well until 128 tiles.

Precision Medicine will rely on frequent genomic analysis, especially for patients undergoing cancer treatments or suffering from rare diseases. Sequence alignment is invoked in multiple stages of the genomic analysis pipeline. Recent projects have introduced accelerators, GenAx and Darwin, for 2nd and 3rd generation sequencers respectively. In this work, we improve upon the GenAx design by increasing its parallelism and reducing its memory bandwidth demands. This is achieved with a combination of hardware and software innovations. We first integrate in-cache operators from prior work into the GenAx memory hierarchy; we then augment the in-cache peripheral circuit to support additional new operators. We then re-structure the sequence alignment algorithm to (i) leverage the many in-cache operators, (ii) exploit the common case in genomic datasets, (iii) use Bloom Filters to reduce futile accesses, and (iv) maximize data reuse within a re-organized memory hierarchy. While the baseline GenAx accelerator processes a batch of reads in 194 seconds while nearly saturating the 153.6 GB/s memory bandwidth, the proposed GenCache architecture processes the same batch of reads in 37 seconds at an improved energy efficiency of 8.6×, while demanding 20 GB/s average memory bandwidth. Our hardware and software techniques thus interact synergistically to target both memory and compute bottlenecks, while not affecting the outputs of the application. We show that the basic principles in GenCache can also be exploited by 3rd generation sequence aligners.

Intel's SGX offers state-of-the-art security features, including confidentiality, integrity, and authentication (CIA) when accessing sensitive pages in memory. Sensitive pages are placed in an Enclave Page Cache (EPC) within the physical memory before they can be accessed by the processor. To control the overheads imposed by CIA guarantees, the EPC operates with a limited capacity (currently 128 MB). Because of this limited EPC size, sensitive pages must be frequently swapped between EPC and non-EPC regions in memory. A page swap is expensive (about 40K cycles) because it requires an OS system call, page copying, updates to integrity trees and metadata, etc. Our analysis shows that the paging overhead can slow the system on average by 5×, and other studies have reported even higher slowdowns for memory-intensive workloads. The paging overhead can be reduced by growing the size of the EPC to match the size of physical memory, while allowing the EPC to also accommodate non-sensitive pages. However, at least two important problems must be addressed to enable this growth in EPC: (i) the depth of the integrity tree and its cacheability must be improved to keep memory bandwidth overheads in check, (ii) the space overheads of integrity verification (tree and MACs) must be reduced. We achieve both goals by introducing a variable arity unified tree (VAULT) organization that is more compact and has lower depth. We further reduce the space overheads with techniques that combine MAC sharing and compression. With simulations, we show that the combination of our techniques can address most inefficiencies in SGX memory access and improve overall performance by 3.7×, relative to an SGX baseline, while incurring a memory capacity over-head of only 4.7%.

As more critical applications move to the cloud, there is a pressing need to provide privacy guarantees for data and computation. While cloud infrastructures are vulnerable to a variety of attacks, in this work, we focus on an attack model where an untrusted cloud operator has physical access to the server and can monitor the signals emerging from the processor socket. Even if data packets are encrypted, the sequence of addresses touched by the program serves as an information side channel. To eliminate this side channel, Oblivious RAM constructs have been investigated for decades, but continue to pose large overheads. In this work, we make the case that ORAM overheads can be significantly reduced by moving some ORAM functionality into the memory system. We first design a secure DIMM (or SDIMM) that uses commodity low-cost memory and an ASIC as a secure buffer chip. We then design two new ORAM protocols that leverage SDIMMs to reduce bandwidth, latency, and energy per ORAM access. In both protocols, each SDIMM is responsible for part of the ORAM tree. Each SDIMM performs a number of ORAM operations that are not visible to the main memory channel. By having many SDIMMs in the system, we are able to achieve highly parallel ORAM operations. The main memory channel uses its bandwidth primarily to service blocks requested by the CPU, and to perform a small subset of the many shuffle operations required by conventional ORAM. The new protocols guarantee the same obliviousness properties as Path ORAM. On a set of memory-intensive workloads, our two new ORAM protocols – Independent ORAM and Split ORAM – are able to improve performance by 1.9x and energy by 2.55x, compared to Freecursive ORAM.

We live in an advent of specialized tasks ranging from graphics, to networking and graph processing, to machine learning and more. While hardware accelerators cater to mainstream demands, general purpose units will always be challenged to run new software. Introspective Computing focuses on building a feedback mechanism to tune dynamic hardware features in real-time. Unlike most prior work, our study is done completely on a real system using hardware resources tunable in most modern Intel processors.

In recent years, multiple neuromorphic architectures have been designed to execute cognitive applications that deal with image and speech analysis. These architectures have followed one of two approaches. One class of architectures is based on machine learning with artificial neural networks. A second class is focused on emulating biology with spiking neuron models, in an attempt to eventually approach the brain's accuracy and energy efficiency. A prominent example of the second class is IBM's TrueNorth processor that can execute large spiking networks on a low-power tiled architecture, and achieve high accuracy on a variety of tasks. However, as we show in this work, there are many inefficiencies in the TrueNorth design. We propose a new architecture, INXS, for spiking neural networks that improves upon the computational efficiency and energy efficiency of the TrueNorth design by 3,129× and 10× respectively. The architecture uses memristor crossbars to compute the effects of input spikes on several neurons in parallel. Digital units are then used to update neuron state. We show that the parallelism offered by crossbars is critical in achieving high throughput and energy efficiency.

Historically, server designers have opted for simple memory systems by picking one of a few commoditized DDR memory products. We are already witnessing a major upheaval in the off-chip memory hierarchy, with the introduction of many new memory products—buffer-on-board, LRDIMM, HMC, HBM, and NVMs, to name a few. Given the plethora of choices, it is expected that different vendors will adopt different strategies for their high-capacity memory systems, often deviating from DDR standards and/or integrating new functionality within memory systems. These strategies will likely differ in their choice of interconnect and topology, with a significant fraction of memory energy being dissipated in I/O and data movement. To make the case for memory interconnect specialization, this paper makes three contributions. First, we design a tool that carefully models I/O power in the memory system, explores the design space, and gives the user the ability to define new types of memory interconnects/topologies. The tool is validated against SPICE models, and is integrated into version 7 of the popular CACTI package. Our analysis with the tool shows that several design parameters have a significant impact on I/O power. We then use the tool to help craft novel specialized memory system channels. We introduce a new relay-on-board chip that partitions a DDR channel into multiple cascaded channels. We show that this simple change to the channel topology can improve performance by 22% for DDR DRAM and lower cost by up to 65% for DDR DRAM. This new architecture does not require any changes to DIMMs, and it efficiently supports hybrid DRAM/NVM systems. Finally, as an example of a more disruptive architecture, we design a custom DIMM and parallel bus that moves away from the DDR3/DDR4 standards. To reduce energy and improve performance, the baseline data channel is split into three narrow parallel channels and the on-DIMM interconnects are operated at a lower frequency. In addition, this allows us to design a two-tier error protection strategy that reduces data transfers on the interconnect. This architecture yields a performance improvement of 18% and a memory power reduction of 23%. The cascaded channel and narrow channel architectures serve as case studies for the new tool and show the potential for benefit from re-organizing basic memory interconnects.

Future systems dealing with big-data workloads will be severely constrained by the high performance and energy penalty imposed by data movement. This penalty can be reduced by storing datasets in DRAM or NVM main memory in compressed formats. Prior compressed memory systems have required significant changes to the operating system, thus limiting commercial viability. The first contribution of this paper is to integrate compression metadata with ECC metadata so that the compressed memory system can be implemented entirely in hardware with no OS involvement. We show that in such a system, read operations are unable to exploit the benefits of compression because the compressibility of the block is not known beforehand. To address this problem, we introduce a compressibility predictor that yields an accuracy of 97%. We also introduce a new data mapping policy that is able to maximize read/write parallelism and NVM endurance, when dealing with compressed blocks. Combined, our proposals are able to eliminate OS involvement and improve performance by 7% (DRAM) and 8% (NVM), and system energy by 12% (DRAM) and 14% (NVM), relative to an uncompressed memory system.

Continued process scaling must overcome several manufacturing challenges. At the same time, industry is exploring many new memory technologies that require new manufacturing processes. In such challenging fabrication regimes, parameter variation (PV) and yield will be important problems. While many recent bodies of work have targeted PV in processors, few have targeted PV in the memory system. Mitigation techniques have either focused on refresh, or have focused on inter-die variation. In this work, with empirical measurements, we first show that PV and specifically intra-die PV is indeed a real phenomenon in modern DRAM chips. We show that this intra-die PV can impact timing parameters for different banks within a DRAM chip. In response to growing PV, memory timing parameters will likely be set very conservatively to accommodate the worst case. To overcome these worst-case limitations, we propose the design of a reconfigurable memory module that detects PV in the field and organizes the memory system into fast/slow regions. This requires changes to the memory controller and to buffer chips on DIMMs. Further, OS migration policies can move frequently accessed pages to the fast regions. This overall approach not only improves performance and energy, it also provides a configurable platform for systems that can tolerate errors or approximation. The proposed system yields an average performance improvement of 12.6% in DRAM systems, and 25.5% in NVM systems.

A number of recent efforts have attempted to design accelerators for popular machine learning algorithms, such as those involving convolutional and deep neural networks (CNNs and DNNs). These algorithms typically involve a large number of multiply-accumulate (dot-product) operations. A recent project, DaDianNao, adopts a near data processing approach, where a specialized neural functional unit performs all the digital arithmetic operations and receives input weights from adjacent eDRAM banks. This work explores an in-situ processing approach, where memristor crossbar arrays not only store input weights, but are also used to perform dot-product operations in an analog manner. While the use of crossbar memory as an analog dot-product engine is well known, no prior work has designed or characterized a full-fledged accelerator based on crossbars. In particular, our work makes the following contributions: (i) We design a pipelined architecture, with some crossbars dedicated for each neural network layer, and eDRAM buffers that aggregate data between pipeline stages. (ii) We define new data encoding techniques that are amenable to analog computations and that can reduce the high overheads of analog-to-digital conversion (ADC). (iii) We define the many supporting digital components required in an analog CNN accelerator and carry out a design space exploration to identify the best balance of memristor storage/compute, ADCs, and eDRAM storage on a chip. On a suite of CNN and DNN workloads, the proposed ISAAC architecture yields improvements of 14.8×, 5.5×, and 7.5× in throughput, energy, and computational density (respectively), relative to the state-of-the-art DaDianNao architecture.

In future memory systems, some regions of memory will be periodically unavailable to the processor. In DRAM systems, this may happen because a rank is busy performing refresh. In non-volatile memory systems, this may happen because a rank is busy draining long-latency writes. Unfortunately, such service interruptions can introduce stalls in all running threads. This is because the operating system spreads the pages of a thread across all memory ranks. Therefore, the probability of a thread accessing data in an unavailable rank is high. This is a performance artifact that has previously not been carefully analyzed. To reduce these stalls, we propose a simple page coloring mechanism that tries to minimize the number of ranks over which a thread's pages are spread. This approach ensures that a service interruption in a single rank only stalls a subset of threads; non-stalled threads even have the potential to run faster at this time because of reduced bus contention. Our analysis shows that this approach is more effective than recent hardware-based mechanisms to deal with such service interruptions. For example, when dealing with service interruptions because of DRAM refresh, the proposed page coloring approach yields an execution time that is 15% lower than the best competing hardware approach.

The scalability of DRAM faces challenges from increasing power consumption and the difficulty of building high aspect ratio capacitors. Consequently, emerging memory technologies including Phase Change Memory (PCM), Spin-Transfer Torque RAM (STT-RAM), and Resistive RAM (ReRAM) are being actively pursued as replacements for DRAM memory. Among these candidates, ReRAM has superior characteristics such as high density, low write energy, and high endurance, making it a very attractive cost-efficient alternative to DRAM. In this paper, we present a comprehensive study of ReRAM-based memory systems. ReRAM's high density comes from its unique crossbar architecture where some peripheral circuits are laid below multiple layers of ReRAM cells. A crossbar architecture introduces special constraints on operating voltages, write latency, and array size. The access latency of a crossbar is a function of the data patterns involved in a write operation. These combined with ReRAM's exponential relationship between its write voltage and switching latency provide opportunities for architectural optimizations. This paper makes several key contributions. First, we study the crossbar architecture and describe trade-offs involving voltage drop, write latency, and data pattern. We then analyze microarchitectural enhancements such as double-sided ground biasing and multiphase reset operations to improve write performance. At the architecture level, a simple compression based data encoding scheme is proposed to further bring down the latency. As the compressibility of a block varies based on its content, write latency is not uniform across blocks. To mitigate the impact of slow writes on performance, we propose and evaluate a novel scheduling policy that makes writing decisions based on latency and activity of a bank. The experimental results show that our architecture improves the performance of a system using ReRAM-based main memory by about 44% over a conservative baseline and 14% over an aggressive baseline on average, and has less than 10% performance degradation compared to an ideal DRAM-only system.

A large fraction of MapReduce execution time is spent processing the Map phase, and a large fraction of Map phase execution time is spent sorting the intermediate key-value pairs generated by the Map function. Sorting accelerators can achieve high performance and low power because they lack the overheads of sorting implementations on general purpose hardware, such as instruction fetch and decode. We find that sorting accelerators are a good match for 3D-stacked Near Data Processing (NDP) because their sorting throughput is so high that it saturates the memory bandwidth available in other memory organizations. The increased sorting performance and low power requirement of fixed-function hardware lead to very high Map phase performance and energy efficiency, reducing Map phase execution time by up to 92%, and reducing energy consumption by up to 91%. We further find that sorting accelerators in a less exotic form of NDP outperform more expensive forms of 3D-stacked NDP without accelerators. We also implement the accelerator on an FPGA to validate our claims.

Several memory vendors are pursuing different kinds of memory cells that can offer high density, non-volatility, high performance, and high endurance. There are several on-going efforts to architect main memory systems with these new NVMs that can compete with traditional DRAM systems. Each NVM has peculiarities that require new microarchitectures and protocols for memory access. In this work, we focus on memristor technology and the sneak currents inherent in memristor crossbar arrays. A read in state-of-the-art designs requires two consecutive reads; the first measures background sneak currents that can be factored out of the current measurement in the second read. This paper introduces a mechanism to reuse the background sneak current measurement for subsequent reads from the same column, thus introducing "open-column" semantics for memristor array access. We also examine a number of data mapping policies that allow the system to balance parallelism and locality. We conclude that on average, it is better to prioritize locality; our best design yields a 20% improvement in read latency and a 26% memory power reduction, relative to the state-of-the-art memristor baseline.

Trusted applications frequently execute in tandem with untrusted applications on personal devices and in cloud environments. Since these co-scheduled applications share hardware resources, the latencies encountered by the untrusted application betray information about whether the trusted applications are accessing shared resources or not. Prior studies have shown that such information leaks can be used by the untrusted application to decipher keys or launch covert-channel attacks. Prior work has also proposed techniques to eliminate information leakage in various shared resources. The best known solution to eliminate information leakage in the memory system incurs high performance penalties. This work develops a comprehensive approach to eliminate timing channels in the memory controller that has two key elements: (i) We shape the memory access behavior of each thread so that it has an unchanging memory access pattern. (ii) We show how efficient memory access pipelines can be constructed to process the resulting memory accesses without introducing any resource conflicts. We mathematically show that the proposed system yields zero information leakage. We then show that various page mapping policies can impact the throughput of our secure memory system. We also introduce techniques to re-order requests from different threads to boost performance without leaking information. Our best solution offers throughput that is 27% lower than that of an optimized non-secure baseline, and that is 69% higher than the best known competing scheme.

Prior work in hardware prefetching has focused mostly on either predicting regular streams with uniform strides, or predicting irregular access patterns at the cost of large hardware structures. This paper introduces the Variable Length Delta Prefetcher (VLDP), which builds up delta histories between successive cache line misses within physical pages, and then uses these histories to predict the order of cache line misses in new pages. One of VLDP's distinguishing features is its use of multiple prediction tables, each of which stores predictions based on a different length of input history. For example, the first prediction table takes as input only the single most recent delta between cache misses within a page, and attempts to predict the next cache miss in that page. The second prediction table takes as input a sequence of the two most recent deltas between cache misses within a page, and also attempts to predict the next cache miss in that page, and so on with additional tables. Longer histories generally yield more accurate predictions, so VLDP prefers to make predictions based on the longest history table that has a matching entry.

Memory latency is a major factor in limiting CPU performance, and prefetching is a well-known method for hiding memory latency. Overly aggressive prefetching can waste scarce resources such as memory bandwidth and cache capacity, limiting or even hurting performance. It is therefore important to employ prefetching mechanisms that use these resources prudently, while still prefetching required data in a timely manner. In this work, we propose a new mechanism to determine at run-time the appropriate prefetching mechanism for the currently executing program, called Sandbox Prefetching. Sandbox Prefetching evaluates simple, aggressive offset prefetchers at run-time by adding the prefetch address to a Bloom filter, rather than actually fetching the data into the cache. Subsequent cache accesses are tested against the contents of the Bloom filter to see if the aggressive prefetcher under evaluation could have accurately prefetched the data, while simultaneously testing for the existence of prefetchable streams. Real prefetches are performed when the accuracy of evaluated prefetchers exceeds a threshold. This method combines the ideas of global pattern confirmation and immediate prefetching action to achieve high performance. Sandbox Prefetching improves performance across the tested workloads by 47.6% compared to not using any prefetching, and by 18.7% compared to the Feedback Directed Prefetching technique. Performance is also improved by 1.4% compared to the Access Map Pattern Matching Prefetcher, while incurring considerably less logic and storage overheads.

Memory compression has been proposed and deployed in the past to grow the capacity of a memory system and reduce page fault rates. Compression also has secondary benefits: it can reduce energy and bandwidth demands. However, most prior mechanisms have been designed to focus on the capacity metric and few prior works have attempted to explicitly reduce energy or bandwidth. Further, mechanisms that focus on the capacity metric also require complex logic to locate the requested data in memory. In this paper, we design a highly simple compressed memory architecture that does not target the capacity metric. Instead, it focuses on complexity, energy, bandwidth, and reliability. It relies on rank subsetting and a careful placement of compressed data and metadata to achieve these benefits. Further, the space made available via compression is used to boost other metrics - the space can be used to implement stronger error correction codes or energy-efficient data encodings. The best performing MemZip configuration yields a 45% performance improvement and 57% memory energy reduction, compared to an uncompressed non-sub-ranked baseline. Another energy-optimized configuration yields a 29.8% performance improvement and a 79% memory energy reduction, relative to the same baseline.

While Processing-in-Memory has been investigated for decades, it has not been embraced commercially. A number of emerging technologies have renewed interest in this topic. In particular, the emergence of 3D stacking and the imminent release of Micron's Hybrid Memory Cube device have made it more practical to move computation near memory. However, the literature is missing a detailed analysis of a killer application that can leverage a Near Data Computing (NDC) architecture. This paper focuses on in-memory MapReduce workloads that are commercially important and are especially suitable for NDC because of their embarrassing parallelism and largely localized memory accesses. The NDC architecture incorporates several simple processing cores on a separate, non-memory die in a 3D-stacked memory package; these cores can perform Map operations with efficient memory access and without hitting the bandwidth wall. This paper describes and evaluates a number of key elements necessary in realizing efficient NDC operation: (i) low-EPI cores, (ii) long daisy chains of memory devices, (iii) the dynamic activation of cores and SerDes links. Compared to a baseline that is heavily optimized for MapReduce execution, the NDC design yields up to 15X reduction in execution time and 18X reduction in system energy.

Memory controllers in modern GPUs aggressively reorder requests for high bandwidth usage, often interleaving requests from different warps. This leads to high variance in the latency of different requests issued by the threads of a warp. Since a warp in a SIMT architecture can proceed only when all of its memory requests are returned by memory, such latency divergence causes significant slowdown when running irregular GPGPU applications. To solve this issue, we propose memory scheduling mechanisms that avoid inter-warp interference in the DRAM system to reduce the average memory stall latency experienced by warps. We further reduce latency divergence through mechanisms that coordinate scheduling decisions across multiple independent memory channels. Finally we show that carefully orchestrating the memory scheduling policy can achieve low average latency for warps, without compromising bandwidth utilization. Our combined scheme yields a 10.1% performance improvement for irregular GPGPU workloads relative to a throughput-optimized GPU memory controller.

Large web-scale applications typically use a distributed platform, like clusters of commodity servers, to achieve scalable and low-cost processing. The Map-Reduce framework and its open-source implementation, Hadoop, is commonly used to program these applications. Since these applications scale well with an increased number of servers, the cluster size is an important parameter. Cluster size however is constrained by power consumption. In this paper we present a system that uses low-power CPUs to increase the cluster size in a fixed power budget. Using low-power CPUs leads to the situation where the majority of a server's power is now consumed by the I/O sub-system. To overcome this, we develop a virtualized I/O sub-system where multiple servers share I/O resources. An ASIC based high-bandwidth interconnect fabric, and FPGA based I/O cards implement this virtualized I/O. The resulting system is the first production quality implementation of cluster-in-a-box that uses low-power CPUs. The unique design demonstrates a way to build systems using low-power CPUs, allowing a much larger number of servers in a cluster in the same power envelope. To overcome software inefficiency and increase the utilization of virtualized disk bandwidth, optimizations necessary for the operating system are also discussed. We built hardware based on these ideas and experiments on this system show a 3X average improvement in performance-per-Watt-hour compared to a commodity cluster with the same power budget.

Many of the pins on a modern chip are used for power delivery. If fewer pins were used to supply the same current, the wires and pins used for power delivery would have to carry larger currents over longer distances. This results in an "IR-drop" problem, where some of the voltage is dropped across the long resistive wires making up the power delivery network, and the eventual circuits experience fluctuations in their supplied voltage. The same problem also manifests if the pin count is the same, but the current draw is higher. IR-drop can be especially problematic in 3D DRAM devices because (i) low cost (few pins and TSVs) is a high priority, (ii) 3D-stacking increases current draw within the package without providing proportionate room for more pins, and (iii) TSVs add to the resistance of the power delivery network.

Co-location of applications is a proven technique to improve hardware utilization. Recent advances in virtualization have made co-location of independent applications on shared hardware a common scenario in datacenters. Co-location, while maintaining Quality-of-Service (QoS) for each application is a complex problem that is fast gaining relevance for these datacenters. The problem is exacerbated by the need for effective resource utilization at datacenter scales. In this work, we show that the memory system is a primary bottleneck in many workloads and is a more effective focal point when enforcing QoS. We examine four different memory system levers to enforce QoS: two that have been previously proposed, and two novel levers. We compare the effectiveness of each lever in minimizing power and resource needs, while enforcing QoS guarantees. We also evaluate the effectiveness of combining various levers and show that this combined approach can yield power reductions of up to 28%.

Many memory cell technologies are being considered as possible replacements for DRAM and Flash technologies, both of which are nearing their scaling limits. While these new cells (PCM, STT-RAM, FeRAM, etc.) promise high density, better scaling, and non-volatility, they introduce new challenges. Solutions at the architecture level can help address some of these problems; e.g., prior research has proposed wear-leveling and hard error tolerance mechanisms to overcome the limited write endurance of PCM cells. In this paper, we focus on the soft error problem in PCM, a topic that has received little attention in the architecture community. Soft errors in DRAM memories are typically addressed by having SECDED support and a scrub mechanism. The scrub mechanism scans the memory looking for a single-bit error and corrects it before the line experiences a second uncorrectable error. However, PCM (and other emerging memories) are prone to new sources of soft errors. In particular, multi-level cell (MLC) PCM devices will suffer from resistance drift, that increases the soft error rate and incurs high overheads for the scrub mechanism. This paper is the first to study the design of architectural scrub mechanisms, especially when tailored to the drift phenomenon in MLC PCM. Many of our solutions will also apply to other soft-error prone emerging memories. We first show that scrub overheads can be reduced with support for strong ECC codes and a lightweight error detection operation. We then design different scrub algorithms that can adaptively trade-off soft and hard errors. Using an approach that combines all proposed solutions, our scrub mechanism yields a 96.5% reduction in uncorrectable errors, a 24.4 × decrease in scrub-related writes, and a 37.8% reduction in scrub energy, relative to a basic scrub algorithm used in modern DRAM systems.

Main memory latencies have always been a concern for system performance. Given that reads are on the critical path for CPU progress, reads must be prioritized over writes. However, writes must be eventually processed and they often delay pending reads. In fact, a single channel in the main memory system offers almost no parallelism between reads and writes. This is because a single off-chip memory bus is shared by reads and writes and the direction of the bus has to be explicitly turned around when switching from writes to reads. This is an expensive operation and its cost is amortized by carrying out a burst of writes or reads every time the bus direction is switched. As a result, no reads can be processed while a memory channel is busy servicing writes. This paper proposes a novel mechanism to boost read-write parallelism and perform useful components of read operations even when the memory system is busy performing writes. If some of the banks are busy servicing writes, we start issuing reads to the other idle banks. The results of these reads are stored in a few registers near the memory chip's I/O pads. These results are quickly returned immediately following the bus turnaround. The process is referred to as a Staged Read because it decouples a single read operation into two stages, with the first step being performed in parallel with writes. This innovation can also be viewed as a form of prefetch that is internal to a memory chip. The proposed technique works best when there is bank imbalance in the write stream. We also introduce a write scheduling algorithm that artificially creates bank imbalance and allows useful read operations to be performed during the write drain. Across a suite of memory-intensive workloads, we show that Staged Reads can boost throughput by up to 33% (average 7%) with an average DRAM access latency improvement of 17%, while incurring a very small cost (0.25%) in terms of memory chip area. The throughput improvements are even greater when considering write-intensive workloads (average 11%) or future systems (average 12%).

Memory system reliability is a serious and growing concern in modern servers. Existing chipkill-level memory protection mechanisms suffer from several draw-backs. They activate a large number of chips on every memory access - this increases energy consumption, and reduces performance due to the reduction in rank-level parallelism. Additionally, they increase access granularity, resulting in wasted bandwidth in the absence of sufficient access locality. They also restrict systems to use narrow-I/O ×4 devices, which are known to be less energy-efficient than the wider ×8 DRAM devices. In this paper, we present LOT-ECC, a localized and multi-tiered protection scheme that attempts to solve these problems. We separate error detection and error correction functionality, and employ simple checksum and parity codes effectively to provide strong fault-tolerance, while simultaneously simplifying implementation. Data and codes are localized to the same DRAM row to improve access efficiency. We use system firmware to store correction codes in DRAM data memory and modify the memory controller to handle data mapping. We thus build an effective fault-tolerance mechanism that provides strong reliability guarantees, activates as few chips as possible (reducing power consumption by up to 44.8% and reducing latency by up to 46.9%), and reduces circuit complexity, all while working with commodity DRAMs and operating systems. Finally, we propose the novel concept of a heterogeneous DIMM that enables the extension of LOT-ECC to ×16 and wider DRAM parts.

The DRAM main memory system in modern servers is largely homogeneous. In recent years, DRAM manufacturers have produced chips with vastly differing latency and energy characteristics. This provides the opportunity to build a heterogeneous main memory system where different parts of the address space can yield different latencies and energy per access. The limited prior work in this area has explored smart placement of pages with high activities. In this paper, we propose a novel alternative to exploit DRAM heterogeneity. We observe that the critical word in a cache line can be easily recognized beforehand and placed in a low-latency region of the main memory. Other non-critical words of the cache line can be placed in a low-energy region. We design an architecture that has low complexity and that can accelerate the transfer of the critical word by tens of cycles. For our benchmark suite, we show an average performance improvement of 12.9% and an accompanying memory energy reduction of 15%.

Modern processors are experiencing interleaved memory access streams from different threads/cores, reducing the spatial locality that is seen at the memory controller, making the combined stream appear increasingly random. Traditional methods for exploiting locality at the DRAM level, such as open-page and timer-based policies, become less effective as the number of threads accessing memory increases. Employing closed-page policies in such systems can improve performance but it eliminates any possibility of exploiting locality. In this paper, we build upon the key insight that a history-based predictor that tracks the number of accesses to a given DRAM page is a much better indicator of DRAM locality than timer based policies. We extend prior work to propose a simple Access Based Predictor (ABP) that tracks limited access history at the page level to determine page closure decisions, and does so with much smaller storage overhead than previously proposed policies. We show that ABP, with additional optimizations, can improve system throughput by 12.3% and 21.6% over open and closed-page policies, respectively. The proposed ABP requires 20 KB of storage overhead and is outside the critical path of memory access.

Future scalable multi-core chips are expected to implement a shared last-level cache (LLC) with banks distributed on chip, forcing a core to incur non-uniform access latencies to each bank. Consequently, high performance and energy efficiency depend on whether a thread's data is placed in local or nearby banks. Using compiler and programmer support, we aim to find an alternative solution to existing high-overhead designs. In this paper, we take existing parallel programs written in Pthreads, and show the performance gap between current static mapping schemes, costly migration schemes and idealized static and dynamic best-case scenarios.

It is well-known that memory latency, energy, capacity, bandwidth, and scalability will be critical bottlenecks in future large-scale systems. This paper addresses these problems, focusing on the interface between the compute cores and memory, comprising the physical interconnect and the memory access protocol. For the physical interconnect, we study the prudent use of emerging silicon-photonic technology to reduce energy consumption and improve capacity scaling. We conclude that photonics are effective primarily to improve socket-edge bandwidth by breaking the pin barrier, and for use on heavily utilized links. For the access protocol, we propose a novel packet based interface that relinquishes most of the tight control that the memory controller holds in current systems and allows the memory modules to be more autonomous, improving flexibility and interoperability. The key enabler here is the introduction of a 3D-stacked interface die that allows both these optimizations without modifying commodity memory dies. The interface die handles all conversion between optics and electronics, as well as all low-level memory device control functionality. Communication beyond the interface die is fully electrical, with TSVs between dies and low-swing wires on-die. We show that such an approach results in substantially lowered energy consumption, reduced latency, better scalability to large capacities, and better support for heterogeneity and interoperability.

Modern processors such as Tilera's Tile64, Intel's Nehalem, and AMD's Opteron are migrating memory controllers (MCs) on-chip, while maintaining a large, flat memory address space. This trend to utilize multiple MC's will likely continue and a core or socket will consequently need to route memory requests to the appropriate MC via an inter- or intra-socket interconnect fabric similar to AMD's HyperTransport(TM), or Intel's Quick-Path Interconnect(TM). Such systems are therefore subject to non-uniform memory access (NUMA) latencies because of the time spent traveling to remote MCs. Each MC will act as the gateway to a particular piece of the physical memory. Data placement will therefore become increasingly critical in minimizing memory access latencies.

Snooping and directory-based coherence protocols have become the de facto standard in chip multi-processors, but neither design is without drawbacks. Snooping protocols are not scalable, while directory protocols incur directory storage overhead, frequent indirections, and are more prone to design bugs. In this paper, we propose a novel coherence protocol that greatly reduces the number of coherence operations and falls back on a simple broadcast-based snooping protocol when infrequent coherence is required. This new protocol is based on the premise that most blocks are either private to a core or read-only, and hence, do not require coherence. This will be especially true for future large-scale multi-core machines that will be used to execute message-passing workloads in the HPC domain, or multiple virtual machines for servers. In such systems, it is expected that a very small fraction of blocks will be both shared and frequently written, hence the need to optimize coherence protocols for a new common case. In our new protocol, dubbed SWEL (protocol states are Shared, Written, Exclusivity Level), the L1 cache attempts to store only private or read-only blocks, while shared and written blocks must reside at the shared L2 level. These determinations are made at runtime without software assistance. While accesses to blocks banished from the L1 become more expensive, SWEL can improve throughput because directory indirection is removed for many common write-sharing patterns. Compared to a MESI based directory implementation, we see up to 15% increased performance, a maximum degradation of 2%, and an average performance increase of 2.5% using SWEL and its derivatives. Other advantages of this strategy are reduced protocol complexity (achieved by reducing transient states) and significantly less storage overhead than traditional directory protocols.

Power consumption and DRAM latencies are serious concerns in modern chip-multiprocessor (CMP or multi-core) based compute systems. The management of the DRAM row buffer can significantly impact both power consumption and latency. Modern DRAM systems read data from cell arrays and populate a row buffer as large as 8 KB on a memory request. But only a small fraction of these bits are ever returned back to the CPU. This ends up wasting energy and time to read (and subsequently write back) bits which are used rarely. Traditionally, an open-page policy has been used for uni-processor systems and it has worked well because of spatial and temporal locality in the access stream. In future multi-core processors, the possibly independent access streams of each core are interleaved, thus destroying the available locality and significantly under-utilizing the contents of the row buffer. In this work, we attempt to improve row-buffer utilization for future multi-core systems.

As manycore architectures enable a large number of cores on the die, a key challenge that emerges is the availability of memory bandwidth with conventional DRAM solutions. To address this challenge, integration of large DRAM caches that provide as much as 5× higher bandwidth and as low as 1/3rd of the latency (as compared to conventional DRAM) is very promising. However, organizing and implementing a large DRAM cache is challenging because of two primary tradeoffs: (a) DRAM caches at cache line granularity require too large an on-chip tag area that makes it undesirable and (b) DRAM caches with larger page granularity require too much bandwidth because the miss rate does not reduce enough to overcome the bandwidth increase. In this paper, we propose CHOP (Caching HOt Pages) in DRAM caches to address these challenges. We study several filter-based DRAM caching techniques: (a) a filter cache (CHOP-FC) that profiles pages and determines the hot subset of pages to allocate into the DRAM cache, (b) a memory-based filter cache (CHOP-MFC) that spills and fills filter state to improve the accuracy and reduce the size of the filter cache and (c) an adaptive DRAM caching technique (CHOP-AFC) to determine when the filter cache should be enabled and disabled for DRAM caching. We conduct detailed simulations with server workloads to show that our filter-based DRAM caching techniques achieve the following: (a) on average over 30% performance improvement over previous solutions, (b) several magnitudes lower area overhead in tag space required for cache-line based DRAM caches, (c) significantly lower memory bandwidth consumption as compared to page-granular DRAM caches.

It is expected that future on-chip networks for many-core processors will impose huge overheads in terms of energy, delay, complexity, verification effort, and area. There is a common belief that the bandwidth necessary for future applications can only be provided by employing packet-switched networks with complex routers and a scalable directory-based coherence protocol. We posit that such a scheme might likely be overkill in a well designed system in addition to being expensive in terms of power because of a large number of power-hungry routers. We show that bus-based networks with snooping protocols can significantly lower energy consumption and simplify network/protocol design and verification, with no loss in performance. We achieve these characteristics by dividing the chip into multiple segments, each having its own broadcast bus, with these buses further connected by a central bus. This helps eliminate expensive routers, but suffers from the energy overhead of long wires. We propose the use of multiple Bloom filters to effectively track data presence in the cache and restrict bus broadcasts to a subset of segments, significantly reducing energy consumption. We further show that the use of OS page coloring helps maximize locality and improves the effectiveness of the Bloom filters. We also employ low-swing wiring to further reduce the energy overheads of the links. Performance can also be improved at relatively low costs by utilizing more of the abundant metal budgets on-chip and employing multiple address-interleaved buses rather than multiple routers. Thus, with the combination of all the above innovations, we extend the scalability of buses and believe that buses can be a viable and attractive option for future on-chip networks. We show energy reductions of up to 31X on average compared to many state-of-the-art packet switched networks.

DRAM vendors have traditionally optimized the cost-per-bit metric, often making design decisions that incur energy penalties. A prime example is the overfetch feature in DRAM, where a single request activates thousands of bit-lines in many DRAM chips, only to return a single cache line to the CPU. The focus on cost-per-bit is questionable in modern-day servers where operating costs can easily exceed the purchase cost. Modern technology trends are also placing very different demands on the memory system: (i)queuing delays are a significant component of memory access time, (ii) there is a high energy premium for the level of reliability expected for business-critical computing, and (iii) the memory access stream emerging from multi-core systems exhibits limited locality. All of these trends necessitate an overhaul of DRAM architecture, even if it means a slight compromise in the cost-per-bit metric.

In the past ten years, computer architecture has seen a paradigm shift from emphasizing single thread performance to energy efficient, throughput oriented, chip multiprocessors. Several studies have suggested that it may be worthwhile to off-load execution of the operating system (OS) to one or more of these cores, or reconfigure hardware during OS execution. To be effective, these techniques must balance the cost of off-loading or re-configuration, versus the potential benefits, which are typically unknown at decision time. These decision points are typically implemented by manually instrumenting a few OS routines (out of hundreds). Such a preliminary research effort cannot be sustained across several operating systems and hardware configurations. We argue that decisions made in software are often sub-optimal because they are expensive in terms of run-time overhead and because applications vary in their use of OS features. We propose that these decision mechanisms should be supported through a hardware based OS run-length predictor, that removes the onus from OS developers. Our final design results in a 95% prediction accuracy for OS intensive applications, while requiring only 2 KB of storage.

Modern processors dedicate more than half their chip area to large L2 and L3 caches and these caches contribute significantly to the total processor power. A large cache is typically split into multiple banks and these banks are either connected through a bus (uniform cache access - UCA) or an on-chip network (non-uniform cache access - NUCA). Irrespective of the cache model (NUCA or UCA), the complex interconnects that must be navigated within large caches are found to be the dominant part of cache access power. While there have been a number of proposals to minimize energy consumption in the inter-bank network, very little attention has been paid to the optimization of intra-bank network power that contributes more than 50% of the total cache dynamic power in large cache banks. In this work we study various mechanisms that introduce low-swing wires inside cache banks as energy saving measures. We propose a novel non-uniform power access design, which when coupled with simple architectural mechanisms, provides the best power-performance tradeoff. The proposed mechanisms reduce cache bank energy by 42% while incurring a minor 1% drop in performance.

In future multi-cores, large amounts of delay and power will be spent accessing data in large L2/L3 caches. It has been recently shown that OS-based page coloring allows a non-uniform cache architecture (NUCA) to provide low latencies and not be hindered by complex data search mechanisms. In this work, we extend that concept with mechanisms that dynamically move data within caches. The key innovation is the use of a shadow address space to allow hardware control of data placement in the L2 cache while being largely transparent to the user application and off-chip world. These mechanisms allow the hardware and OS to dynamically manage cache capacity per thread as well as optimize placement of data shared by multiple threads. We show an average IPC improvement of 10-20% for multi-programmed workloads with capacity allocation policies and an average IPC improvement of 8% for multi-threaded workloads with policies for shared page placement.

Cache hierarchies in future many-core processors are expected to grow in size and contribute a large fraction of overall processor power and performance. In this paper, we postulate a 3D chip design that stacks SRAM and DRAM upon processing cores and employs OS-based page coloring to minimize horizontal communication of cache data. We then propose a heterogeneous reconfigurable cache design that takes advantage of the high density of DRAM and the superior power/delay characteristics of SRAM to efficiently meet the working set demands of each individual core. Finally, we analyze the communication patterns for such a processor and show that a tree topology is an ideal fit that significantly reduces the power and latency requirements of the on-chip network. The above proposals are synergistic: each proposal is made more compelling because of its combination with the other innovations described in this paper. The proposed reconfigurable cache model improves performance by up to 19% along with 48% savings in network power.

In a hardware transactional memory system with lazy versioning and lazy conflict detection, the process of transaction commit can emerge as a bottleneck. This is especially true for a large-scale distributed memory system where multiple transactions may attempt to commit simultaneously and coordination is required before allowing commits to proceed in parallel. In this paper, we propose novel algorithms to implement commit that are more scalable in terms of delay and are free of deadlocks/livelocks. We show that these algorithms have similarities with the token cache coherence concept and leverage these similarities to extend the algorithms to handle message loss and starvation scenarios. The proposed algorithms improve upon the state-of-the-art by yielding up to a 7X reduction in commit delay and up to a 48X reduction in network messages for commit. These translate into overall performance improvements of up to 66% (for synthetic workloads with average transaction length of 200 cycles), 35% (for average transaction length of 1000 cycles), and 8% (for average transaction length of 4000 cycles). For a small group of multi-threaded programs with frequent transaction commits, improvements of up to 8% were observed for a 32-node simulation.

The ever increasing sizes of on-chip caches and the growing domination of wire delay necessitate significant changes to cache hierarchy design methodologies. Many recent proposals advocate splitting the cache into a large number of banks and employing a network-on-chip (NoC) to allow fast access to nearby banks (referred to as Non-Uniform Cache Architectures--NUCA). Most studies on NUCA organizations have assumed a generic NoC and focused on logical policies for cache block placement, movement, and search. Since wire/router delay and power are major limiting factors in modern processors, this work focuses on interconnect design and its influence on NUCA performance and power. We extend the widely-used CACTI cache modeling tool to take network design parameters into account. With these overheads appropriately accounted for, the optimal cache organization is typically very different from that assumed in prior NUCA studies. To alleviate the interconnect delay bottleneck, we propose novel cache access optimizations that introduce heterogeneity within the inter-bank network. The careful consideration of interconnect choices for a large cache results in a 51% performance improvement over a baseline generic NoC and the introduction of heterogeneity within the network yields an additional 11-15% performance improvement.

Aggressive technology scaling over the years has helped improve processor performance but has caused a reduction in processor reliability. Shrinking transistor sizes and lower supply voltages have increased the vulnerability of computer systems towards transient faults. An increase in within-die and die-to-die parameter variations has also led to a greater number of dynamic timing errors. A potential solution to mitigate the impact of such errors is redundancy via an in-order checker processor. Emerging 3D chip technology promises increased processor performance as well as reduced power consumption because of shorter on-chip wires. In this paper, we leverage the "snap-on" functionality provided by 3D integration and propose implementing the redundant checker processor on a second die. This allows manufacturers to easily create a family of "reliable processors" without significantly impacting the cost or performance for customers that care less about reliability. We comprehensively evaluate design choices for this second die, including the effects of L2 cache organization, deep pipelining, and frequency. An interesting feature made possible by 3D integration is the incorporation of heterogeneous process technologies within a single chip. We evaluate the possibility of providing redundancy with an older process technology, an unexplored and especially compelling application of die heterogeneity. We show that with the most pessimistic assumptions, the overhead of the second die can be as high as either a 7degC temperature increase or a 8% performance loss. However, with the use of an older process, this overhead can be reduced to a 3degC temperature increase or a 4% performance loss, while also providing higher error resilience.

A significant part of future microprocessor real estate will be dedicated to 12 or 13 caches. These on-chip caches will heavily impact processor performance, power dissipation, and thermal management strategies. There are a number of interconnect design considerations that influence power/performance/area characteristics of large caches, such as wire models (width/spacing/repeaters), signaling strategy (RC/differential/transmission), router design, etc. Yet, to date, there exists no analytical tool that takes all of these parameters into account to carry out a design space exploration for large caches and estimate an optimal organization. In this work, we implement two major extensions to the CACTI cache modeling tool that focus on interconnect design for a large cache. First, we add the ability to model different types of wires, such as RC-based wires with different power/delay characteristics and differential low-swing buses. Second, we add the ability to model non-uniform cache access (NUCA). We not only adopt state-of-the-art design space exploration strategies for NUCA, we also enhance this exploration by considering on-chip network contention and a wider spectrum of wiring and routing choices. We present a validation analysis of the new tool (to be released as CACTI 6.0) and present a case study to showcase how the tool can improve architecture research methodologies.

Noise and radiation-induced soft errors (transient faults) in computer systems have increased significantly over the last few years and are expected to increase even more as we move towards smaller transistor sizes and lower supply voltages. Fault detection and recovery can be achieved through redundancy. The emergence of chip multiprocessors (CMPs) makes it possible to execute redundant threads on a chip and provide relatively low-cost reliability. State-of-the-art implementations execute two copies of the same program as two threads (redundant multithreading), either on the same or on separate processor cores in a CMP, and periodically check results. While this solution has favorable performance and reliability properties, every redundant instruction flows through a high-frequency complex outof- order pipeline, thereby incurring a high power consumption penalty. This paper proposes mechanisms that attempt to provide reliability at a modest power and complexity cost. When executing a redundant thread, the trailing thread benefits from the information produced by the leading thread. We take advantage of this property and comprehensively study different strategies to reduce the power overhead of the trailing core in a CMP. These strategies include dynamic frequency scaling, in-order execution, and parallelization of the trailing thread.

Improvements in semiconductor technology have made it possible to include multiple processor cores on a single die. Chip Multi-Processors (CMP) are an attractive choice for future billion transistor architectures due to their low design complexity, high clock frequency, and high throughput. In a typical CMP architecture, the L2 cache is shared by multiple cores and data coherence is maintained among private L1s. Coherence operations entail frequent communication over global on-chip wires. In future technologies, communication between different L1s will have a significant impact on overall processor performance and power consumption. On-chip wires can be designed to have different latency, bandwidth, and energy properties. Likewise, coherence protocol messages have different latency and bandwidth needs. We propose an interconnect composed of wires with varying latency, bandwidth, and energy characteristics, and advocate intelligently mapping coherence operations to the appropriate wires. In this paper, we present a comprehensive list of techniques that allow coherence protocols to exploit a heterogeneous interconnect and evaluate a subset of these techniques to show their performance and power-efficiency potential. Most of the proposed techniques can be implemented with a minimum complexity overhead.

The ever increasing demand for high clock speeds and the desire to exploit abundant transistor budgets have resulted in alarming increases in processor power dissipation. Partitioned (or clustered) architectures have been proposed in recent years to address scalability concerns in future billion-transistor microprocessors. Our analysis shows that increasing processor resources in a clustered architecture results in a linear increase in power consumption, while providing diminishing improvements in single-thread performance. To preserve high performance to power ratios, we claim that the power consumption of additional resources should be in proportion to the performance improvements they yield. Hence, in this paper, we propose the implementation of heterogeneous clusters that have varying delay and power characteristics. A cluster's performance and power characteristic is tuned by scaling its frequency and novel policies dynamically assign frequencies to clusters, while attempting to either meet a fixed power budget or minimize a metric such as Energy /spl times/ Delay/sup 2/ (ED/sup 2/). By increasing resources in a power-efficient manner, we observe an 11% improvement in ED/sup 2/ and a 22.4% average reduction in peak temperature, when compared to a processor with homogeneous units. Our proposed processor model also provides strategies to handle thermal emergencies that have a relatively low impact on performance.

Future high-performance billion-transistor processors are likely to employ partitioned architectures to achieve high clock speeds, high parallelism, low design complexity, and low power. In such architectures, inter-partition communication over global wires has a significant impact on overall processor performance and power consumption. VLSI techniques allow a variety of wire implementations, but these wire properties have previously never been exposed to the microarchitecture. This paper advocates global wire management at the microarchitecture level and proposes a heterogeneous interconnect that is comprised of wires with varying latency, bandwidth, and energy characteristics. We propose and evaluate microarchitectural techniques that can exploit such a heterogeneous interconnect to improve performance and reduce energy consumption. These techniques include a novel cache pipeline design, the identification of narrow bit-width operands, the classification of non-critical data, and the detection of interconnect load imbalance. For a dynamically scheduled partitioned architecture, our results demonstrate that the proposed innovations result in up to 11% reductions in overall processor ED/sup 2/, compared to a baseline processor that employs a homogeneous interconnect.

The growing dominance of wire delays at future technology points renders a microprocessor communication-bound. Clustered microarchitectures allow most dependence chains to execute without being affected by long on-chip wire latencies. They also allow faster clock speeds and reduce design complexity, thereby emerging as a popular design choice for future microprocessors. However, a centralized data cache threatens to be the primary bottle-neck in highly clustered systems. The paper attempts to identify the most complexity-effective approach to alleviate this bottleneck. While decentralized cache organizations have been proposed, they introduce excessive logic and wiring complexity. The paper evaluates if the performance gains of a decentralized cache are worth the increase in complexity. We also introduce and evaluate the behavior of Cluster Prefetch - the forwarding of data values to a cluster through accurate address prediction. Our results show that the success of this technique depends on accurate speculation across unresolved stores. The technique applies for a wide class of processor models and most importantly, it allows high performance even while employing a simple centralized data cache. We conclude that address prediction holds more promise for future wire-delay-limited processors than decentralized cache organizations.

Clustered microarchitectures are an attractive alternative to large monolithic superscalar designs due to their potential for higher clock rates in the face of increasingly wire-delay-constrained process technologies. As increasing transistor counts allow an increase in the number of clusters, thereby allowing more aggressive use of instruction-level parallelism (ILP), the inter-cluster communication increases as data values get spread across a wider area. As a result of the emergence of this trade-off between communication and parallelism, a subset of the total on-chip clusters is optimal for performance. To match the hardware to the application's needs, we use a robust algorithm to dynamically tune the clustered architecture. The algorithm, which is based on program metrics gathered at periodic intervals, achieves an 11% performance improvement on average over the best statically defined architecture. We also show that the use of additional hardware and reconfiguration at basic block boundaries can achieve average improvements of 15%. Our results demonstrate that reconfiguration provides an effective solution to the communication and parallelism trade-off inherent in the communication-bound processors of the future.

The widespread use of repeaters in long wires creates the possibility of dynamically sizing regular on-chip structures. We present a tunable cache and translation lookaside buffer (TLB) hierarchy that leverages repeater insertion to dynamically trade off size for speed and power consumption on a per-application phase basis using a novel configuration management algorithm. In comparison to a conventional design that is fixed at a single design point targeted to the average application, the dynamically tunable cache and TLB hierarchy can be tailored to the needs of each application phase. The configuration algorithm dynamically detects phase changes and selects a configuration based on the application's ability to tolerate different hit and miss latencies in order to improve the memory energy-delay product. We evaluate the performance and energy consumption of our approach and project the effects of technology scaling trends on our design.

As clock frequency increases and feature size decreases, clock distribution and wire delays present a growing challenge to the designers of singly-clocked, globally synchronous systems. We describe an alternative approach, which we call a multiple clock domain (MCD) processor, in which the chip is divided into several clock domains, within which independent voltage and frequency scaling can be performed. Boundaries between domains are chosen to exploit existing queues, thereby minimizing inter-domain synchronization costs. We propose four clock domains, corresponding to the front end , integer units, floating point units, and load-store units. We evaluate this design using a simulation infrastructure based on SimpleScalar and Wattch. In an attempt to quantify potential energy savings independent of any particular on-line control strategy, we use off-line analysis of traces from a single-speed run of each of our benchmark applications to identify profitable reconfiguration points for a subsequent dynamic scaling run. Using applications from the MediaBench, Olden, and SPEC2000 benchmark suites, we obtain an average energy-delay product improvement of 20% with MCD compared to a modest 3% savings from voltage scaling a single clock and voltage system.

Modern superscalar processors use wide instruction issue widths and out-of-order execution in order to increase instruction-level parallelism (ILP). Because instructions must be committed in order so as to guarantee precise exceptions, increasing ILP implies increasing the sizes of structures such as the register file, issue queue, and reorder buffer. Simultaneously, cycle time constraints limit the sizes of these structures, resulting in conflicting design requirements.

Dynamic superscalar processors execute multiple instructions out-of-order by looking for independent operations within a large window. The number of physical registers within the processor has a direct impact on the size of this window as most in-flight instructions require a new physical register at dispatch. A large multi-ported register file helps improve the instruction-level parallelism (ILP), but may have a detrimental effect on clock speed, especially in future wire-limited technologies. In this paper, we propose a register file organization that reduces register file size and port requirements for a given amount of ILP. We use a two-level register file organization to reduce register file size requirements, and a banked organization to reduce port requirements. We demonstrate empirically that the resulting register file organizations have reduced latency and (in the case of the banked organization) energy requirements for similar instructions per cycle (IPC) performance and improved instructions per second (IPS) performance in comparison to a conventional monolithic register file. The choice of organization is dependent on design goals.

Conventional microarchitectures choose a single memory hierarchy design point targeted at the average application. In this paper, we propose a cache and TLB layout and design that leverages repeater insertion to provide dynamic low-cost configurability trading off size and speed on a per application phase basis. A novel configuration management algorithm dynamically detects phase changes and reacts to an application's hit and miss intolerance in order to improve memory hierarchy performance while taking energy consumption into consideration. When applied to a two-level cache and TLB hierarchy at 0.1 /spl mu/m technology, the result is an average 15% reduction in cycles per instruction (CPI), corresponding to an average 27% reduction in memory-CPI, across a broad class of applications compared to the best conventional two-level hierarchy of comparable size. Projecting to sub-.1 /spl mu/m technology design considerations that call for a three-level conventional cache hierarchy for performance reasons, we demonstrate that a configurable L2/L3 cache hierarchy coupled with a conventional LI results in an average 43% reduction in memory hierarchy energy in addition to improved performance.