Achieving peak throughput on modern CPUs requires maximizing the use of single-instruction, multiple-data (SIMD) or vector compute units. Single-program, multiple-data (SPMD) programming models are an effective way to use high-level programming languages to target these ISAs. Unfortunately, many SPMD frameworks have evolved to have either overly-restrictive language specifications or under-specified programming models, and this has slowed the widescale adoption of SPMD-style programming. This paper introduces Parsimony (PARallel SIMd), a SPMD programming approach built with semantics designed to be compatible with multiple languages and to cleanly integrate into the standard optimizing compiler toolchains for those languages. We first explain the Parsimony programming model semantics and how they enable a standalone compiler IR-to-IR pass that can perform vectorization independently of other passes, improving the language and toolchain compatibility of SPMD programming. We then demonstrate a LLVM prototype of the Parsimony approach that matches the performance of ispc, a popular but more restrictive SPMD approach, and achieves 97% of the performance of hand-written AVX-512 SIMD intrinsics on over 70 benchmarks ported from the Simd Library. We finally discuss where Parsimony has exposed parts of existing language and compiler flows where slight improvements could further enable improved SPMD program vectorization.

Recent studies have shown that using fine-grained peer-to-peer (P2P) stores to communicate among devices in multi-GPU systems is a promising path to achieve strong performance scaling. In many irregular applications, such as graph algorithms and sparse linear algebra, small sub-cache line (4-32B) stores arise naturally when using the P2P paradigm. This is particularly problematic in multi-GPU systems because inter-GPU interconnects are optimized for bulk transfers rather than small operations. As a consequence, application developers either resort to complex programming techniques to work around this small transfer inefficiency or fall back to bulk inter-GPU DMA transfers that have limited performance scalability. We propose FinePack, a set of limited I/O interconnect and GPU hardware enhancements that enable small peer-to-peer stores to achieve interconnect efficiency that rivals bulk transfers while maintaining the simplicity of a peer-to-peer memory access programming model. Exploiting the GPU’s weak memory model, FinePack dynamically coalesces and compresses small writes into a larger I/O message that reduces link-level protocol overhead. FinePack is fully transparent to software and requires no changes to the GPU’s virtual memory system. We evaluate FinePack on a system comprising 4 Volta GPUs on a PCIe 4.0 interconnect to show FinePack improves interconnect efficiency for small peer-to-peer stores by 3×. This results in 4-GPU strong scaling performance 1.4× better than traditional DMA based multi-GPU programming and comes within 71% of the maximum achievable strong scaling performance.

Irregular, memory-intensive applications often incur high translation lookaside buffer (TLB) miss rates that result in significant address translation overheads. Employing huge pages is an effective way to reduce these overheads, however in real systems the number of available huge pages can be limited when system memory is nearly full and/or fragmented. Thus, huge pages must be used selectively to back application memory. This work demonstrates that choosing memory regions that incur the most TLB misses for huge page promotion best reduces address translation overheads. We call these regions High reUse TLB-sensitive data (HUBs). Unlike prior work which relies on expensive per-page software counters to identify promotion regions, we propose new architectural support to identify these regions dynamically at application runtime.

As GPUs scale their low-precision matrix math throughput to boost deep learning (DL) performance, they upset the balance between math throughput and memory system capabilities. We demonstrate that a converged GPU design trying to address diverging architectural requirements between FP32 (or larger)-based HPC and FP16 (or smaller)-based DL workloads results in sub-optimal configurations for either of the application domains. We argue that a C omposable O n- PA ckage GPU (COPA-GPU) architecture to provide domain-specialized GPU products is the most practical solution to these diverging requirements. A COPA-GPU leverages multi-chip-module disaggregation to support maximal design reuse, along with memory system specialization per application domain. We show how a COPA-GPU enables DL-specialized products by modular augmentation of the baseline GPU architecture with up to 4× higher off-die bandwidth, 32× larger on-package cache, and 2.3× higher DRAM bandwidth and capacity, while conveniently supporting scaled-down HPC-oriented designs. This work explores the microarchitectural design necessary to enable composable GPUs and evaluates the benefits composability can provide to HPC, DL training, and DL inference. We show that when compared to a converged GPU design, a DL-optimized COPA-GPU featuring a combination of 16× larger cache capacity and 1.6× higher DRAM bandwidth scales per-GPU training and inference performance by 31% and 35%, respectively, and reduces the number of GPU instances by 50% in scale-out training scenarios.

The demands of high-performance computing (HPC) and machine learning (ML) workloads have resulted in the rapid architectural evolution of GPUs over the last decade. The growing memory footprint and diversity of data types in these workloads has required GPUs to embrace micro-architectural heterogeneity and increased memory system sophistication to scale performance. Effective simulation of new architectural features early in the design cycle enables quick and effective exploration of design trade-offs across this increasingly diverse set of workloads. This work provides a retrospective on the design and development of NVArchSim (NVAS), an architectural simulator used within NVIDIA to design and evaluate features that are difficult to appraise using other methodologies due to workload type, size, complexity, or lack of modeling flexibility. We argue that overly precise and/or overly slow architectural models hamper an architect's ability to evaluate new features within a reasonable time frame, hurting productivity. Because of its speed, NVAS is being used to trace and evaluate hundreds of HPC and state-of-the-art ML workloads on single-GPU or multi-GPU systems. By adding component fidelity only when necessary to improve system-level modeling accuracy, NVAS delivers simulation speed orders of magnitude higher than most publicly available GPU simulators while retaining high levels of accuracy and simulation flexibility. Building trustworthy high-level simulation platforms is a difficult exercise in balance and compromise; we share our experiences to help and encourage those in academia who take on the challenge of building GPU simulation platforms.

Despite continuing research into inter-GPU communication mechanisms, extracting performance from multi-GPU systems remains a significant challenge. Inter-GPU communication via bulk DMA-based transfers exposes data transfer latency on the GPU’s critical execution path because these large transfers are logically interleaved between compute kernels. Conversely, fine-grained peer-to-peer memory accesses during kernel execution lead to memory stalls that can exceed the GPUs’ ability to cover these operations via multi-threading. Worse yet, these sub-cacheline transfers are highly inefficient on current inter-GPU interconnects. To remedy these issues, we propose PROACT, a system enabling remote memory transfers with the programmability and pipeline advantages of peer-to-peer stores, while achieving interconnect efficiency that rivals bulk DMA transfers. Combining compile-time instrumentation with fine-grain tracking of data block readiness within each GPU, PROACT enables interconnect-friendly data transfers while hiding the transfer latency via pipelining during kernel execution. This work describes both hardware and software implementations of PROACT and demonstrates the effectiveness of a PROACT software prototype on three generations of GPU hardware and interconnects. Achieving near-ideal interconnect efficiency, PROACT realizes a mean speedup of 3.0× over single-GPU performance for 4-GPU systems, capturing 83% of available performance opportunity. On a 16-GPU NVIDIA DGX-2 system, we demonstrate an 11.0× average strong-scaling speedup over single-GPU performance, 5.3× better than a bulk DMA-based approach.

Suboptimal management of memory and bandwidth is one of the primary causes of low performance on systems comprising multiple GPUs. Existing memory management solutions like Unified Memory (UM) offer simplified programming but come at the cost of performance: applications can even exhibit slowdown with increasing GPU count due to their inability to leverage system resources effectively. To solve this challenge, we propose GPS, a HW/SW multi-GPU memory management technique that efficiently orchestrates inter-GPU communication using proactive data transfers. GPS offers the programmability advantage of multi-GPU shared memory with the performance of GPU-local memory. To enable this, GPS automatically tracks the data accesses performed by each GPU, maintains duplicate physical replicas of shared regions in each GPU’s local memory, and pushes updates to the replicas in all consumer GPUs. GPS is compatible within the existing NVIDIA GPU memory consistency model but takes full advantage of its relaxed nature to deliver high performance. We evaluate GPS in the context of a 4-GPU system with varying interconnects and show that GPS achieves an average speedup of 3.0 × relative to the performance of a single GPU, outperforming the next best available multi-GPU memory management technique by 2.3 × on average. In a 16-GPU system, using a future PCIe 6.0 interconnect, we demonstrate a 7.9 × average strong scaling speedup over single-GPU performance, capturing 80% of the available opportunity.

Prior work on GPU cache coherence has shown that simple hardware-or software-based protocols can be more than sufficient. However, in recent years, features such as multi-chip modules have added deeper hierarchy and non-uniformity into GPU memory systems. GPU programming models have chosen to expose this non-uniformity directly to the end user through scoped memory consistency models. As a result, there is room to improve upon earlier coherence protocols that were designed only for flat single-GPU hierarchies and/or simpler memory consistency models. In this paper, we propose HMG, a cache coherence protocol designed for forward-looking multi-GPU systems. HMG strikes a balance between simplicity and performance: it uses a readily-implementable VI-like protocol to track coherence states, but it tracks sharers using a hierarchical scheme optimized for mitigating the bandwidth limitations of inter-GPU links. HMG leverages the novel scoped, non-multi-copy-atomic properties of modern GPU memory models, and it avoids the overheads of invalidation acknowledgments and transient states that were needed to support prior GPU memory models. On a 4-GPU system, HMG improves performance over a software-controlled, bulk invalidation-based coherence mechanism by 26% and over a non-hierarchical hardware cache coherence protocol by 18%, thereby achieving 97% of the performance of an idealized caching system.

GPUs accelerate high-throughput applications, which require orders-of-magnitude higher memory bandwidth than traditional CPU-only systems. However, the capacity of such high-bandwidth memory tends to be relatively small. Buddy Compression is an architecture that makes novel use of compression to utilize a larger buddy-memory from the host or disaggregated memory, effectively increasing the memory capacity of the GPU. Buddy Compression splits each compressed 128B memory-entry between the high-bandwidth GPU memory and a slower-but-larger buddy memory such that compressible memory-entries are accessed completely from GPU memory, while incompressible entries source some of their data from off-GPU memory. With Buddy Compression, compressibility changes never result in expensive page movement or re-allocation. Buddy Compression achieves on average 1.9× effective GPU memory expansion for representative HPC applications and 1.5× for deep learning training, performing within 2% of an unrealistic system with no memory limit. This makes Buddy Compression attractive for performance-conscious developers that require additional GPU memory capacity.

Recent work has shown that building GPUs with hundreds of SMs in a single monolithic chip will not be practical due to slowing growth in transistor density, low chip yields, and photoreticle limitations. To maintain performance scalability, proposals exist to aggregate discrete GPUs into a larger virtual GPU and decompose a single GPU into multiple-chip-modules with increased aggregate die area. These approaches introduce non-uniform memory access (NUMA) effects and lead to decreased performance and energy-efficiency if not managed appropriately. To overcome these effects, we propose a holistic Locality-Aware Data Management (LADM) system designed to operate on massive logical GPUs composed of multiple discrete devices, which are themselves composed of chiplets. LADM has three key components: a threadblock-centric index analysis, a runtime system that performs data placement and threadblock scheduling, and an adaptive cache insertion policy. The runtime combines information from the static analysis with topology information to proactively optimize data placement, threadblock scheduling, and remote data caching, minimizing off-chip traffic. Compared to state-of-the-art multi-GPU scheduling, LADM reduces inter-chip memory traffic by 4× and improves system performance by 1.8× on a future multi-GPU system.

Software-controlled heterogeneous memory systems have the potential to increase the performance and cost efficiency of computing systems. However they can only deliver on this promise if supported by efficient page management policies and mechanisms within the operating system (OS). Current OS implementations do not support efficient tiering of data between heterogeneous memories. Instead, they rely on expensive offlining of memory or swapping data to disk as a means of profiling and migrating hot or cold data between memory nodes. They also leave numerous optimizations on the table; for example, multi-threaded hardware is not leveraged to maximize page migration throughput, resulting in up to 95% under-utilization of available memory bandwidth. To remedy these shortcomings, we propose and implement a general purpose OS-integrated multi-level memory management system that reuses current OS page tracking structures to tier pages directly between memories with no additional monitoring overhead. We augment this system with four additional optimizations: native support for transparent huge page migration, multi-threaded migration of a page, concurrent migration of multiple pages, and symmetric exchange of pages. Combined, these optimizations dramatically reduce kernel software overheads and improve raw page migration throughput over 15×. Implemented in Linux and evaluated on x86, Power, and ARM64 systems, our OS support for heterogeneous memories improves application performance 40% over baseline Linux for a suite of real-world memory-intensive workloads utilizing a multi-level disaggregated memory system.

As Moore's law slows down, GPUs must pivot towards multi-module designs to continue scaling performance at historical rates. Prior work on multi-module GPUs has focused on performance, while largely ignoring the issue of energy efficiency. In this work, we propose a new metric for GPU efficiency called EDP Scaling Efficiency that quantifies the effects of both strong performance scaling and overall energy efficiency in these designs. To enable this analysis, we develop a novel top-down GPU energy estimation framework that is accurate within 10% of a recent GPU design. Being decoupled from granular GPU microarchitectural details, the framework is appropriate for energy efficiency studies in future GPUs. Using this model in conjunction with performance simulation, we show that the dominating factor influencing the energy efficiency of GPUs over the next decade is GPUmodule (GPM) idle time. Furthermore, neither inter-module interconnect energy, nor GPM microarchitectural design is expected to play a key role in this regard. We demonstrate that multi-module GPUs are on a trajectory to become 2× less energy efficient than current monolithic designs; a significant issue for data centers which are already energy constrained. Finally, we show that architects must be willing to spend more (not less) energy to enable higher bandwidth inter-GPM connections, because counter-intuitively, this additional energy expenditure can reduce total GPU energy consumption by as much as 45%, providing a path to energy efficient strong scaling in the future.

Virtual memory (VM) eases programming effort but can suffer from high address translation overheads. Architects have traditionally coped by increasing Translation Lookaside Buffer (TLB) capacity; this approach, however, requires considerable hardware resources. One promising alternative is to rely on software-generated translation contiguity to compress page translation encodings within the TLB. To enable this, operating systems (OSes) have to assign spatially-adjacent groups of physical frames to contiguous groups of virtual pages, as doing so allows compression or coalescing of these contiguous translations in hardware. Unfortunately, modern OSes do not currently guarantee translation contiguity in many real-world scenarios; as systems remain online for long periods of time, their memory can and does become fragmented.

Binary instrumentation frameworks are widely used to implement profilers, performance evaluation, error checking, and bug detection tools. While dynamic binary instrumentation tools such as PIN and DynamoRio are supported on CPUs, GPU architectures currently only have limited support for similar capabilities through static compile-time tools, which prohibits instrumentation of dynamically loaded libraries that are foundations for modern high-performance applications. This work presents NVBit, a fast, dynamic, and portable, binary instrumentation framework, that allows users to write instrumentation tools in CUDA/C/C++ and selectively apply that functionality to pre-compiled binaries and libraries executing on NVIDIA GPUs. Using dynamic recompilation at the SASS level, NVBit analyzes GPU kernel register requirements to generate efficient ABI compliant instrumented code without requiring the tool developer to have detailed knowledge of the underlying GPU architecture. NVBit allows basic-block instrumentation, multiple function injections to the same location, inspection of all ISA visible state, dynamic selection of instrumented or uninstrumented code, permanent modification of register state, source code correlation, and instruction removal. NVBit supports all recent NVIDIA GPU architecture families including Kepler, Maxwell, Pascal and Volta and works on any pre-compiled CUDA, OpenACC, OpenCL, or CUDA-Fortran application.

Historically, improvement in GPU performance has been tightly coupled with transistor scaling. As Moore's Law slows down, performance of single GPUs may ultimately plateau. To continue GPU performance scaling, multiple GPUs can be connected using system-level interconnects. However, limited inter-GPU interconnect bandwidth (e.g., 64GB/s) can hurt multi-GPU performance when there are frequent remote GPU memory accesses. Traditional GPUs rely on page migration to service the memory accesses from local memory instead. Page migration fails when the page is simultaneously shared between multiple GPUs in the system. As such, recent proposals enhance the software runtime system to replicate read-only shared pages in local memory. Unfortunately, such practice fails when there are frequent remote memory accesses to read-write shared pages. To address this problem, recent proposals cache remote shared data in the GPU last-level-cache (LLC). Unfortunately, remote data caching also fails when the shared-data working-set exceeds the available GPU LLC size. This paper conducts a combined performance analysis of state-of-the-art software and hardware mechanisms to improve NUMA performance of multi-GPU systems. Our evaluations on a 4-node multi-GPU system reveal that the combination of work scheduling, page placement, page migration, page replication, and caching remote data still incurs a 47% slowdown relative to an ideal NUMA-GPU system. This is because the shared memory footprint tends to be significantly larger than the GPU LLC size and can not be replicated by software because the shared footprint has read-write property. Thus, we show that existing NUMA-aware software solutions require hardware support to address the NUMA bandwidth bottleneck. We propose Caching Remote Data in Video Memory (CARVE), a hardware mechanism that stores recently accessed remote shared data in a dedicated region of the GPU memory. CARVE outperforms state-of-the-art NUMA mechanisms and is within 6% the performance of an ideal NUMA-GPU system. A design space analysis on supporting cache coherence is also investigated. Overall, we show that dedicating only 3% of GPU memory eliminates NUMA bandwidth bottlenecks while incurring negligible performance overheads due to the reduced GPU memory capacity.

Historically, improvements in GPU-based high performance computing have been tightly coupled to transistor scaling. As Moore's law slows down, and the number of transistors per die no longer grows at historical rates, the performance curve of single monolithic GPUs will ultimately plateau. However, the need for higher performing GPUs continues to exist in many domains. To address this need, in this paper we demonstrate that package-level integration of multiple GPU modules to build larger logical GPUs can enable continuous performance scaling beyond Moore's law. Specifically, we propose partitioning GPUs into easily manufacturable basic GPU Modules (GPMs), and integrating them on package using high bandwidth and power efficient signaling technologies. We lay out the details and evaluate the feasibility of a basic Multi-Chip-Module GPU (MCM-GPU) design. We then propose three architectural optimizations that significantly improve GPM data locality and minimize the sensitivity on inter-GPM bandwidth. Our evaluation shows that the optimized MCM-GPU achieves 22.8% speedup and 5x inter-GPM bandwidth reduction when compared to the basic MCM-GPU architecture. Most importantly, the optimized MCM-GPU design is 45.5% faster than the largest implementable monolithic GPU, and performs within 10% of a hypothetical (and unbuildable) monolithic GPU. Lastly we show that our optimized MCM-GPU is 26.8% faster than an equally equipped Multi-GPU system with the same total number of SMs and DRAM bandwidth.

GPUs achieve high throughput and power efficiency by employing many small single instruction multiple thread (SIMT) cores. To minimize scheduling logic and performance variance they utilize a uniform memory system and leverage strong data parallelism exposed via the programming model. With Moore's law slowing, for GPUs to continue scaling performance (which largely depends on SIMT core count) they are likely to embrace multi-socket designs where transistors are more readily available. However when moving to such designs, maintaining the illusion of a uniform memory system is increasingly difficult. In this work we investigate multi-socket non-uniform memory access (NUMA) GPU designs and show that significant changes are needed to both the GPU interconnect and cache architectures to achieve performance scalability. We show that application phase effects can be exploited allowing GPU sockets to dynamically optimize their individual interconnect and cache policies, minimizing the impact of NUMA effects. Our NUMA-aware GPU outperforms a single GPU by 1.5×, 2.3×, and 3.2× while achieving 89%, 84%, and 76% of theoretical application scalability in 2, 4, and 8 sockets designs respectively. Implementable today, NUMA-aware multi-socket GPUs may be a promising candidate for scaling GPU performance beyond a single socket.

Cache coherence is ubiquitous in shared memory multiprocessors because it provides a simple, high performance memory abstraction to programmers. Recent work suggests extending hardware cache coherence between CPUs and GPUs to help support programming models with tightly coordinated sharing between CPU and GPU threads. However, implementing hardware cache coherence is particularly challenging in systems with discrete CPUs and GPUs that may not be produced by a single vendor. Instead, we propose, selective caching, wherein we disallow GPU caching of any memory that would require coherence updates to propagate between the CPU and GPU, thereby decoupling the GPU from vendor-specific CPU coherence protocols. We propose several architectural improvements to offset the performance penalty of selective caching: aggressive request coalescing, CPU-side coherent caching for GPU-uncacheable requests, and a CPU-GPU interconnect optimization to support variable-size transfers. Moreover, current GPU workloads access many read-only memory pages; we exploit this property to allow promiscuous GPU caching of these pages, relying on page-level protection, rather than hardware cache coherence, to ensure correctness. These optimizations bring a selective caching GPU implementation to within 93% of a hardware cache-coherent implementation without the need to integrate CPUs and GPUs under a single hardware coherence protocol.

Despite industrial investment in both on-die GPUs and next generation interconnects, the highest performing parallel accelerators shipping today continue to be discrete GPUs. Connected via PCIe, these GPUs utilize their own privately managed physical memory that is optimized for high bandwidth. These separate memories force GPU programmers to manage the movement of data between the CPU and GPU, in addition to the on-chip GPU memory hierarchy. To simplify this process, GPU vendors are developing software runtimes that automatically page memory in and out of the GPU on-demand, reducing programmer effort and enabling computation across datasets that exceed the GPU memory capacity. Because this memory migration occurs over a high latency and low bandwidth link (compared to GPU memory), these software runtimes may result in significant performance penalties. In this work, we explore the features needed in GPU hardware and software to close the performance gap of GPU paged memory versus legacy programmer directed memory management. Without modifying the GPU execution pipeline, we show it is possible to largely hide the performance overheads of GPU paged memory, converting an average 2× slowdown into a 12% speedup when compared to programmer directed transfers. Additionally, we examine the performance impact that GPU memory oversubscription has on application run times, enabling application designers to make informed decisions on how to shard their datasets across hosts and GPU instances.

Systems from smartphones to supercomputers are increasingly heterogeneous, being composed of both CPUs and GPUs. To maximize cost and energy efficiency, these systems will increasingly use globally-addressable heterogeneous memory systems, making choices about memory page placement critical to performance. In this work we show that current page placement policies are not sufficient to maximize GPU performance in these heterogeneous memory systems. We propose two new page placement policies that improve GPU performance: one application agnostic and one using application profile information. Our application agnostic policy, bandwidth-aware (BW-AWARE) placement, maximizes GPU throughput by balancing page placement across the memories based on the aggregate memory bandwidth available in a system. Our simulation-based results show that BW-AWARE placement outperforms the existing Linux INTERLEAVE and LOCAL policies by 35% and 18% on average for GPU compute workloads. We build upon BW-AWARE placement by developing a compiler-based profiling mechanism that provides programmers with information about GPU application data structure access patterns. Combining this information with simple program-annotated hints about memory placement, our hint-based page placement approach performs within 90% of oracular page placement on average, largely mitigating the need for costly dynamic page tracking and migration.

Historically, GPU-based HPC applications have had a substantial memory bandwidth advantage over CPU-based workloads due to using GDDR rather than DDR memory. However, past GPUs required a restricted programming model where application data was allocated up front and explicitly copied into GPU memory before launching a GPU kernel by the programmer. Recently, GPUs have eased this requirement and now can employ on-demand software page migration between CPU and GPU memory to obviate explicit copying. In the near future, CC-NUMA GPU-CPU systems will appear where software page migration is an optional choice and hardware cache-coherence can also support the GPU accessing CPU memory directly. In this work, we describe the trade-offs and considerations in relying on hardware cache-coherence mechanisms versus using software page migration to optimize the performance of memory-intensive GPU workloads. We show that page migration decisions based on page access frequency alone are a poor solution and that a broader solution using virtual address-based program locality to enable aggressive memory prefetching combined with bandwidth balancing is required to maximize performance. We present a software runtime system requiring minimal hardware support that, on average, outperforms CC-NUMA-based accesses by 1.95 ×, performs 6% better than the legacy CPU to GPU memcpy regime by intelligently using both CPU and GPU memory bandwidth, and comes within 28% of oracular page placement, all while maintaining the relaxed memory semantics of modern GPUs.

To aid application characterization and architecture design space exploration, researchers and engineers have developed a wide range of tools for CPUs, including simulators, profilers, and binary instrumentation tools. With the advent of GPU computing, GPU manufacturers have developed similar tools leveraging hardware profiling and debugging hooks. To date, these tools are largely limited by the fixed menu of options provided by the tool developer and do not offer the user the flexibility to observe or act on events not in the menu. This paper presents SASSI (NVIDIA assembly code "SASS" Instrumentor), a low-level assembly-language instrumentation tool for GPUs. Like CPU binary instrumentation tools, SASSI allows a user to specify instructions at which to inject user-provided instrumentation code. These facilities allow strategic placement of counters and code into GPU assembly code to collect user-directed, fine-grained statistics at hardware speeds. SASSI instrumentation is inherently parallel, leveraging the concurrency of the underlying hardware. In addition to the details of SASSI, this paper provides four case studies that show how SASSI can be used to characterize applications and explore the architecture design space along the dimensions of instruction control flow, memory systems, value similarity, and resilience.

Modern scientific discovery is driven by an insatiable demand for computing performance. The HPC community is targeting development of supercomputers able to sustain 1 ExaFlops by the year 2020 and power consumption is the primary obstacle to achieving this goal. A combination of architectural improvements, circuit design, and manufacturing technologies must provide over a 20× improvement in energy efficiency. In this paper, we present some of the progress NVIDIA Research is making toward the design of Exascale systems by tailoring features to address the scaling challenges of performance and energy efficiency. We evaluate several architectural concepts for a set of HPC applications demonstrating expected energy efficiency improvements resulting from circuit and packaging innovations such as low-voltage SRAM, low-energy signalling, and on-package memory. Finally, we discuss the scaling of these features with respect to future process technologies and provide power and performance projections for our Exascale research architecture.

The IO performance of storage devices has accelerated from hundreds of IOPS five years ago, to hundreds of thousands of IOPS today, and tens of millions of IOPS projected in five years. This sharp evolution is primarily due to the introduction of NAND-flash devices and their data parallel design. In this work, we demonstrate that the block layer within the operating system, originally designed to handle thousands of IOPS, has become a bottleneck to overall storage system performance, specially on the high NUMA-factor processors systems that are becoming commonplace. We describe the design of a next generation block layer that is capable of handling tens of millions of IOPS on a multi-core system equipped with a single storage device. Our experiments show that our design scales graciously with the number of cores, even on NUMA systems with multiple sockets.

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.

Over the last twenty years the interfaces for accessing persistent storage within a computer system have remained essentially unchanged. Simply put, seek, read and write have defined the fundamental operations that can be performed against storage devices. These three interfaces have endured because the devices within storage subsystems have not fundamentally changed since the invention of magnetic disks. Non-volatile (flash) memory (NVM) has recently become a viable enterprise grade storage medium. Initial implementations of NVM storage devices have chosen to export these same disk-based seek/read/write interfaces because they provide compatibility for legacy applications. We propose there is a new class of higher order storage primitives beyond simple block I/O that high performance solid state storage should support. One such primitive, atomic-write, batches multiple I/O operations into a single logical group that will be persisted as a whole or rolled back upon failure. By moving write-atomicity down the stack into the storage device, it is possible to significantly reduce the amount of work required at the application, filesystem, or operating system layers to guarantee the consistency and integrity of data. In this work we provide a proof of concept implementation of atomic-write on a modern solid state device that leverages the underlying log-based flash translation layer (FTL). We present an example of how database management systems can benefit from atomic-write by modifying the MySQL InnoDB transactional storage engine. Using this new atomic-write primitive we are able to increase system throughput by 33%, improve the 90th percentile transaction response time by 20%, and reduce the volume of data written from MySQL to the storage subsystem by as much as 43% on industry standard benchmarks, while maintaining ACID transaction semantics.

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.

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.

Simulators for digital systems operate at a variety of levels of abstraction varying from detailed analog and switch level modeling of the transistor to cycle based descriptions of entire systems. We propose an even higher level simulator, called ARCS, based on the abstraction of an asynchronous communication event rather than of a clock cycle. Modeling systems at this level allows architectural level exploration of the design space before cycle-level details are available, and also allows the same framework to be used to refine architectural level simulations into more detailed simulations with increasingly fine grained notions of timing. The ARCS simulation framework uses concurrently operating threads in Java with communicating sequential processes (CSP) semantics as a natural expression of communication between concurrent hardware. To avoid synchronization bottlenecks ARCS models time using a communication driven clockwork model which allows for both user configurable runtime viewing of the simulation and post processing of complete simulation timing data.