Transformers have revolutionized AI in natural language processing and computer vision, but their enormous computation and memory demands pose significant challenges for hardware acceleration. In practice, end-to-end throughput is often limited by paged data movement and interconnect bandwidth, not just raw MAC count. This work proposes a unified system–accelerator co-design approach to efficiently accelerate transformer inference by jointly optimizing a novel hardware matrix accelerator and its system integration with paged, streaming dataflows and explicit overlap of compute and transfer. On the hardware side, we introduce MatrixFlow , a loosely‐coupled 16× 16 systolic‐array accelerator featuring a block‐based matrix multiplication method that is page-aligned (4 KB tiles), uses only a small (≈20 KB) on-chip buffer, and runs a pipelined schedule of DMA, compute, and DMA-out to fully utilize interconnect bandwidth, emphasizing standard DMA-driven streaming rather than large on-chip reuse. On the system side, we develop Gem5‐AcceSys , an extension of the gem5 full system simulator allowing exploration of standard interconnects (PCIe) and configurable memory hierarchies including Direct-Memory (DM), Direct-Cache (DC), and Device-Memory (DevMem) modes with SMMU/TLB effects. Through co‐design, MatrixFlow’s novel dataflow and the Gem5‐AcceSys platform are tuned in tandem to alleviate data‐movement bottlenecks without requiring specialized CPU‐instruction‐set modifications. We validate our approach with gem5 simulations on representative transformer models (BERT and ViT) across multiple data types and system setups. Results demonstrate up to 22× speed‐up in end‐to‐end inference over a CPU‐only baseline and performance gains of 5×–8× over state‐of‐the‐art loosely‐ and tightly‐coupled accelerators. Furthermore, we show that a standard PCIe-based host memory design can achieve ∼80% of the performance of on‐device HBM memory. Overall, paged streaming and pipeline overlap , not large local SRAMs, emerge as the most effective knobs for efficient transformer inference under realistic system constraints.

The growing demand for efficient, high-performance processing in machine learning (ML) and image processing has made hardware accelerators, such as GPUs and Data Streaming Accelerators (DSAs), increasingly essential. These accelerators enhance ML and image processing tasks by offloading computation from the CPU to dedicated hardware. These accelerators rely on interconnects for efficient data transfer, making interconnect design crucial for system-level performance. This paper introduces Gem5-AcceSys, an innovative framework for system-level exploration of standard interconnects and configurable memory hierarchies. Using a matrix multiplication accelerator tailored for transformer workloads as a case study, we evaluate PCIe performance across diverse memory types (DDR4, DDR5, GDDR6, HBM2) and configurations, including host-side and device-side memory. Our findings demonstrate that optimized interconnects can achieve up to $80 \%$ of device-side memory performance and, in some scenarios, even surpass it. These results offer actionable insights for system architects, enabling a balanced approach to performance and cost in next-generation accelerator design.

By interfacing computing logic directly to the DRAM banks, bank-level Compute-near-Memory (CnM) architectures promise to mitigate the bottleneck at the memory interconnect. While this computation paradigm heavily reduces the energy requirements for data movement across the system, current solutions fail to co-optimize hardware and software to further increase efficiency. Instead, in this manuscript, we present SideDRAM , a co-designed bank-level CnM architecture to enable massively parallel and energy-efficient computations near DRAM. In contrast with past solutions, we support flexible data typing and heterogeneous quantization, relying on the robustness of workloads to employ small bitwidths, and enable a row-wide access to the banks to exploit parallelism and spatial locality. As a result, SideDRAM integrates (1) software-defined SIMD (SoftSIMD) datapaths, supporting low-energy computing with flexible precision, (2) an interface to the banks based on very wide registers (VWRs), enabling asymmetric data access to both utilize the full DRAM bank bandwidth and leverage data locality at the datapath, and (3) a low-overhead distributed control plane, allowing the efficient handling of variable data typing. We benchmark SideDRAM as a near-DRAM solution by analyzing the area, performance, and energy consumption of an HBM2 CnM channel executing heterogeneously quantized machine learning models. The results show that, compared to the state-of-the-art FIMDRAM design, energy improvements of up to 67% are achieved when a DeiT-S inference is executed with a batch size of 16 under the same area constraints, resulting in energy-delay-area product (EDAP) savings that reach 83%. When comparing to a massively parallel mixed-signal CnM solution, SideDRAM consistently obtains similar performance and better energy efficiency results (geomean of 15× improvement across workloads) at a lower area overhead.

Full-System (FS) simulation is essential for performance evaluation of complete systems that execute complex applications on a complete software stack consisting of an operating system and user applications. Nevertheless, they require careful fine-tuning against real hardware to obtain reliable performance statistics, which can become tedious, error-prone, and time-consuming with typical trial-and-error approaches. We propose a novel, streamlined, component-level calibration methodology to address these shortcomings to validate FS simulation models. Our methodology greatly accelerates the validation process without sacrificing accuracy. It is Instruction Set Architecture (ISA)-agnostic, and can tackle hardware specifications at different levels of detail. We demonstrate its effectiveness by validating FS models against both open-hardware and IP-protected (closed hardware) RISC-V silicon, achieving a mean error of 19%–23% for the SPEC CPU2017 suite in the two cases. We introduce the first open-source RISC-V-based FS-validated simulation models with a complete and replicable methodology.

In recent years, Artificial Intelligence (AI) has seen a remarkable expansion. Traditionally, applications primarily relied on edge computing, closer to data sources. However, today edge computing approaches are outstripped by modern AI demands. In this paper, we introduce a novel framework together with an online decision algorithm based on multi-armed contextual bandits (DroneBandit), for the dynamic allocation of inference tasks between edge and cloud, to increase inference performance. Our test environment consists of a resource-constrained nanodrone equipped with a custom DNN for obstacle detection, able to achieve 80% accuracy and 71% F2-score. DroneBandit runs on the drone and chooses the optimal cutting point for each iteration on the fly. The decision is based on predicted back-end delays (i.e. data transfer and inference time on the cloud) and observed front-end delays (i.e. inference time on the edge). DroneBandit achieves 83% accuracy and 89% Top3 accuracy in predicting the ideal cutting point on simulations. Our framework demonstrates enhanced task allocation adaptability, enabling efficient computation offloading and edge computing reliance in varying network conditions. Our experiments, conducted on a nanodrone with an ARM CPU and GAP8 RISC-V accelerator, incorporate quantization and optimization, showcasing efficient obstacle detection in dynamic scenarios.

In the past decade, computing systems have significantly increased in complexity and power consumption. Nowadays, heterogeneous multi-processor systems-on-chip (MPSoCs) integrate many computing cores. Heterogeneous MPSoCs often comprise general-purpose processors and a variety of accelerators, thus supporting specialized functions for the target application domain to minimize overall energy when executing a specific task. The ensuing architectural design space is, therefore, increasingly multi-dimensional, especially in the light of upcoming 2.5D/3D chiplets integration, which, on one side, allows unprecedented system integration possibilities but, on the other, exacerbates data transfer bottlenecks and affects overall power consumption significantly. To traverse such space in search of high-performance/high-efficiency solutions, we introduce a cross-layer approach combining fast explorations with virtual systems with modular open-hardware design frameworks. This paper showcases how these two approaches effectively cross-fertilize: detailed hardware designs are essential in calibrating performance, power and temperature models, and validating simulation outcomes. Conversely, full system simulation is crucial for projecting the impact of design choices towards complex but energy-efficient heterogeneous multi-processor architectures.

The increasing demand for computing power and the emergence of heterogeneous computing architectures have driven the exploration of innovative techniques to address current limitations in both the compute and memory subsystems. One such solution is the use of Accelerated Processing Units (APUs), processors that incorporate both a central processing unit (CPU) and an integrated graphics processing unit (iGPU). However, the performance of both APU and CPU systems can be significantly hampered by address translation overhead, leading to a decline in overall performance, especially for cache-resident workloads. To address this issue, we propose the introduction of a new intermediate address space (IAS) in both APU and CPU systems. IAS serves as a bridge between virtual address (VA) spaces and physical address (PA) spaces, optimizing the address translation process. In the case of APU systems, our research indicates that the iGPU suffers from significant translation look-aside buffer (TLB) misses in certain workload situations. Using an IAS, we can divide the initial address translation into front- and back-end phases, effectively shifting the bottleneck in address translation from the cache side to the memory controller side, a technique that proves to be effective for cache-resident workloads. Our simulations demonstrate that implementing IAS in the CPU system can boost performance by up to 40% compared to conventional CPU systems. Furthermore, we evaluate the effectiveness of APU systems, comparing the performance of IAS-based systems with traditional systems, showing up to a 185% improvement in APU system performance with our proposed IAS implementation. Furthermore, our analysis indicates that over 90% of TLB misses can be filtered by the cache, and employing a larger cache within the system could potentially result in even greater improvements. The proposed IAS offers a promising and practical solution to enhance the performance of both APU and CPU systems, contributing to state-of-the-art research in the field of computer architecture.

Near-DRAM computing strategies advocate for providing computational capabilities close to where data is stored. Although this paradigm can effectively address the memory-to-processor communication bottleneck, it also presents new challenges: The strict resource constraints in the memory periphery demand careful tailoring of architectural elements. We herein propose a novel framework and methodology to explore compute-near-memory designs that interface to DRAM memory banks, demonstrating the area, energy, and performance tradeoffs subject to the architectural configuration. We exemplify this methodology by conducting two studies on compute-near-bank designs: 1) analyzing the interaction between control and data resources, and 2) exploring the integration of processing units with different DRAM standards. According to our study, the optimal size ratios between instruction and data capacity vary from$2\times $to$4\times $across benchmarks from representative application domains. The retrieved Pareto-optimal solutions from our framework improve state-of-the-art designs, e.g., achieving a 50% performance increase on matrix operations with 15% energy overhead relative to the FIMDRAM design. In addition, the exploration of DRAM shows the interplay between available internal bandwidth, performance, and area overhead. For example, a threefold increase in bandwidth rises performance by 47% across workloads at a 34% extra area cost.

Due to stringent energy and performance constraints, edge AI computing often employs heterogeneous systems that utilize both general-purpose CPUs and accelerators. Analog in-memory computing (AIMC) is a well-known AI inference solution that overcomes computational bottlenecks by performing matrix-vector multiplication operations (MVMs) in constant time. However, the tiles of AIMC-based accelerators are limited by the number of weights they can hold. State-of-the-art research often sizes neural networks to AIMC tiles (or vice-versa), but does not consider cases where AIMC tiles cannot cover the whole network due to lack of tile resources or the network size. In this work, we study the trade-offs of available AIMC tile resources, neural network coverage, AIMC tile proximity to compute resources, and multi-core load balancing techniques. We first perform a study of single-layer performance and energy scalability of AIMC tiles in the two most typical AIMC acceleration targets: dense/fully-connected layers and convolutional layers. This study guides the methodology with which we approach parameter allocation to AIMC tiles in the context of large edge neural networks, both where AIMC tiles are close to the CPU (tightly-coupled) and cannot share resources across the system, and where AIMC tiles are far from the CPU (loosely-coupled) and can employ workload stealing. We explore the performance and energy trends of six modern CNNs using different methods of load balancing for differently-coupled system configurations with variable AIMC tile resources. We show that, by properly distributing workloads, AIMC acceleration can be made highly effective even on under-provisioned systems. As an example, 5.9x speedup and 5.6x energy gains were measured on an 8-core system, for a 41% coverage of neural network parameters.

Computing servers have played a key role in developing and processing emerging compute-intensive applications in recent years. Consolidating multiple virtual machines (VMs) inside one server to run various applications introduces severe competence for limited resources among VMs. Many techniques such as VM scheduling and resource provisioning are proposed to maximize the cost-efficiency of the computing servers while alleviating the performance inference between VMs. However, these management techniques require accurate performance prediction of the application running inside the VM, which is challenging to get in the public cloud due to the black-box nature of the VMs. From this perspective, this paper proposes a novel machine learning-based performance prediction approach for applications running in the cloud. To achieve high-accuracy predictions for black-box VMs, the proposed method first identifies the running application inside the virtual machine. It then selects highly correlated runtime metrics as the input of the machine learning approach to accurately predict the performance level of the cloud application. Experimental results with state-of-the-art cloud benchmarks demonstrate that our proposed method outperforms existing prediction methods by more than 2× in terms of the worst prediction error. In addition, we successfully tackle the challenge of performance prediction for applications with variable workloads by introducing the performance degradation index, which other comparison methods fail to consider. The workflow versatility of the proposed approach has been verified with different modern servers and VM configurations.

Multi-Chiplet architectures are being increasingly adopted to support the design of very large systems in a single package, facilitating the integration of heterogeneous components and improving manufacturing yield. However, chiplet-based solutions have to cope with limited inter-chiplet routing resources, which complicate the design of the data interconnect and the power delivery network. Emerging in-package wireless technology is a promising strategy to address these challenges, as it allows to implement flexible chiplet interconnects while freeing package resources for power supply connections. To assess the capabilities of such an approach and its impact from a full-system perspective, herein we present an exploration of the performance of in-package wireless communication, based on dedicated extensions to the gem5-X simulator. We consider different Medium Access Control (MAC) protocols, as well as applications with different runtime profiles, showcasing that current in-package wireless solutions are competitive with wired chiplet interconnects. Our results show how in-package wireless solutions can outperform wired alternatives when running artificial intelligence workloads, achieving up to a 2.64× speed-up when running deep neural networks (DNNs) on a chiplet-based system with 16 cores distributed in four clusters.

2.5D Systems-on-Package (SoPs) are composed by several chiplets placed on an interposer. They are becoming increasingly popular as they enable easy integration of electronic components in the same package and high fabrication yields. Nevertheless, they introduce a new bottleneck in inter-chiplet communication, which must be routed through the interposer. Such a constraint favors mapping related tasks on computing cores within the same chiplet, leading to thermal hotspots. In-package wireless technology holds promise to reconsider such a position because integrated wireless antennas provide low-latency and high-bandwidth communication paths, thus bypassing the in-terposer bottleneck. Furthermore, in this work, we propose a new task mapping heuristic that leverages in-package wireless technology to improve the thermal behavior of 2.5D SoPs executing complex applications. Combining system simulation and thermal modeling, our results show that we can distribute computation in wireless 2.5D SoPs to reduce peak temperatures by up to 24% through task mapping with a negligible performance impact.

Analog in-memory computing (AIMC) cores offers significant performance and energy benefits for neural network inference with respect to digital logic (e.g., CPUs). AIMCs accelerate matrix-vector multiplications, which dominate these applications' run-time. However, AIMC-centric platforms lack the flexibility of general-purpose systems, as they often have hard-coded data flows and can only support a limited set of processing functions. With the goal of bridging this gap in flexibility, we present a novel system architecture that tightly integrates analog in-memory computing accelerators into multi-core CPUs in general-purpose systems. We developed a powerful gem5-based full system-level simulation framework into the gem5-X simulator, ALPINE, which enables an in-depth characterization of the proposed architecture. ALPINE allows the simulation of the entire computer architecture stack from major hardware components to their interactions with the Linux OS. Within ALPINE, we have defined a custom ISA extension and a software library to facilitate the deployment of inference models. We showcase and analyze a variety of mappings of different neural network types, and demonstrate up to 20.5x/20.8x performance/energy gains with respect to a SIMD-enabled ARM CPU implementation for convolutional neural networks, multi-layer perceptrons, and recurrent neural networks.

Flow cell arrays (FCAs) concurrently provide efficient on-chip liquid cooling and electrochemical power generation. This technology is especially promising for 3-D multiprocessor systems-on-chip (3-D MPSoCs) realized in deeply scaled technologies, which present very challenging power and thermal requirements. Indeed, FCAs effectively improve power delivery network (PDN) performance, particularly if switched capacitor (SC) converters are employed to decouple the flow cells and the systems-on-chip voltages, allowing each to operate at their optimal point. Nonetheless, the design of FCA-based solutions entails nonobvious considerations and tradeoffs, stemming from their dual role in governing both the thermal and power delivery characteristics of 3-D MPSoCs. Showcasing them in this article, we explore multiple FCA design configurations and demonstrate that this technology can decrease the temperature of a heterogeneous 3-D MPSoC by 78 °C, and its total power consumption by 46%, compared to a high-performance cold-plate-based liquid cooling solution. At the same time, FCAs enable up to 90% voltage drop recovery across dies, using SC converters occupying a small fraction of the chip area. Such outcomes provide an opportunity to boost 3-D MPSoC computing performance by increasing the operating frequency of dies. Leveraging these results, we introduce a novel temperature and voltage-aware model-predictive control (MPC) strategy that optimizes power efficiency during runtime. We achieve application-wide speedups of up to 16% on various machine learning (ML), data mining, and other high-performance benchmarks while keeping the 3-D MPSoC temperature below 83 °C and voltage drops below 5%.

Flow Cell Arrays (FCA) technology employs microchannels filled with an electrolytic fluid to concurrently provide cooling and power generation to integrated circuits (ICs). This solution is particularly appealing for Three-Dimensional Multi-Processor Systems-on-Chip (3D MPSoCs) realized in deeply scaled technologies, as their extreme power densities result in significant thermal and voltage supply challenges. FCAs provide them with extra power to boost performance. However, the dual effects of FCAs (cooling and power supply) have conflicting trends leading to a complex interplay between temperature, voltage stability, and performance. In this paper, we explore this trade-off by introducing a novel methodology that controls the operating frequency of computing components and the electrolytic coolant flow rate at run-time. Our strategy enables tangible performance gains while abiding by timing, voltage drop, and temperature constraints. We showcase its benefits by targeting a 4-layer 3D MPSoC, achieving up to 24% increase in the operating frequencies and resulting in application speedups of up to 17%, while reducing the costs related to FCA liquid pumping energy.

Computing servers play a key role in the development and process of emerging compute-intensive applications in recent years. However, they need to operate efficiently from an energy perspective viewpoint, while maximizing the performance and lifetime of the hottest server components (i.e., cores and cache). Previous methods focused on either improving energy efficiency by adopting new hybrid-cache architectures including the resistive random-access memory (RRAM) and static random-access memory (SRAM) at the hardware level, or exploring tradeoffs between lifetime limitation and performance of multicore processors under stable workloads conditions. Therefore, no work has so far proposed a co-optimization method with hybrid-cache-based server architectures for real-life dynamic scenarios taking into account scalability, performance, lifetime reliability, and energy efficiency at the same time. In this article, we first formulate a reliability model for the hybrid-cache architecture to enable precise lifetime reliability management and energy efficiency optimization. We also include the performance and energy overheads of cache switching, and optimize the benefits of hybrid-cache usage for better energy efficiency and performance. Then, we propose a runtime$q$-learning-based reliability management and performance optimization approach for multicore microprocessors with the hybrid-cache architecture, jointly incorporated with a dynamic preemptive priority queue management method to improve the overall tasks’ performance by targeting to respect their end time limits. Experimental results show that our proposed method achieves up to 44% average performance (i.e., tasks execution time) improvement, while maintaining the whole system design lifetime longer than five years, when compared to the latest state-of-the-art energy efficiency optimization and reliability management methods for computing servers.

High-performance computing (HPC) servers aim to meet an increase in the number and complexity of tasks and, consequently, to address the energy efficiency challenge. In addition to energy efficiency, it is essential to manage lifetime limitations of power-hungry components of servers (e.g., cores and cache), hence avoiding server failure before its lifetime period. Traditional approaches focus on either using hybrid caches to reduce the leakage power of traditional static random-access memory (SRAM) cache, and thus increase the energy efficiency, or the tradeoff between the lifetime and performance of multicore processors. However, these approaches fall short in terms of flexibility and applicability for HPC tasks in terms of multiparametric optimization, including quality-of-service (QoS), lifetime reliability, and energy efficiency. As a result, in this article, we propose COCKTAIL, a holistic strategy framework to jointly optimize the energy efficiency of multicore server processors and tasks performance in the HPC context, while guaranteeing the lifetime reliability. First, we analyze the best cache technology among traditional SRAM and resistive random access memory (RRAM), within the context of hybrid cache architectures, to improve the energy efficiency and manage cache endurance limits with respect to tasks requirements. Second, we introduce a novel efficient proactive queue optimization policy to reorder HPC tasks for execution considering their end time and possible reliability effects on the use of the hybrid caches. Third, we present a dynamic model predictive control (MPC)-based reliability management method to maximize task performance, by controlling the frequency, temperature, and target lifetime of the server processor. Our results demonstrate that, while consuming similar energy, COCKTAIL provides up to 60% QoS improvement when compared to latest state-of-the-art energy optimization and reliability management techniques in the HPC context. Moreover, our strategy guarantees a design lifetime longer than five years for the whole HPC system.

Nowadays, deep convolutional neural networks (DCNNs) play a significant role in many application domains, such as computer vision, medical imaging, and image processing. Nonetheless, designing a DCNN, able to defeat the state of the art, is a manual, challenging, and time-consuming task, due to the extremely large design space, as a consequence of a large number of layers and their corresponding hyperparameters. In this work, we address the challenge of performing hyperparameter optimization of DCNNs through a novel multiagent reinforcement learning (MARL)-based approach, eliminating the human effort. In particular, we adapt$Q$-learning and define learning agents per layer to split the design space into independent smaller design subspaces such that each agent fine tunes the hyperparameters of the assigned layer concerning a global reward. Moreover, we provide a novel formation of$Q$-tables along with a new update rule that facilitates agents’ communication. Our MARL-based approach is data driven and able to consider an arbitrary set of design objectives and constraints. We apply our MARL-based solution to different well-known DCNNs, including GoogLeNet, VGG, and U-Net, and various datasets for image classification and semantic segmentation. Our results have shown that compared to the original CNNs, the MARL-based approach can reduce the model size, training time, and inference time by up to, respectively,$83\times $, 52%, and 54% without any degradation in accuracy. Moreover, our approach is very competitive to state-of-the-art neural architecture search methods in terms of the designed CNN accuracy and its number of parameters while significantly reducing the optimization cost.

The increasing power density in modern high-performance multiprocessor System-on-Chip (MPSoC) is fueling a revolution in thermal management. On the one hand, thermal phenomena are becoming a critical concern, making accurate and efficient simulation a necessity. On the other hand, a variety of physically heterogeneous solutions is coming into play: liquid, evaporative, thermoelectric cooling, and more. A new generation of simulators, with unprecedented flexibility, is thus required. In this article, we present 3D-ICE 3.0, the first thermal simulator to allow for accurate nonlinear descriptions of complex and physically heterogeneous heat dissipation systems, while preserving the efficiency of latest compact modeling frameworks at the silicon die level. 3D-ICE 3.0 allows designers to extend the thermal simulator with new heat sink models while simplifying the time-consuming step of model validation. The support for nonlinear dynamic models is included, for instance, to accurately represent variable coolant flows. Our results present validated models of a commercial water heat sink and an air heat sink plus fan that achieve an average error below 1 °C and simulate, respectively, up to$3\times $and$12\times $faster than the real physical phenomena.

Improving the energy efficiency of data centers while guaranteeing Quality of Service (QoS), together with detecting performance variability of servers caused by either hardware or software failures, are two of the major challenges for efficient resource management of large-scale cloud infrastructures. Previous works in the area of dynamic Virtual Machine (VM) consolidation are mostly focused on addressing the energy challenge, but fall short in proposing comprehensive, scalable, and low-overhead approaches that jointly tackle energy efficiency and performance variability. Moreover, they usually assume over-simplistic power models, and fail to accurately consider all the delay and power costs associated with VM migration and host power mode transition. These assumptions are no longer valid in modern servers executing heterogeneous workloads and lead to unrealistic or inefficient results. In this paper, we propose a centralized-distributed low-overhead failure-aware dynamic VM consolidation strategy to minimize energy consumption in large-scale data centers. Our approach selects the most adequate power mode and frequency of each host during runtime using a distributed multi-agent Machine Learning (ML) based strategy, and migrates the VMs accordingly using a centralized heuristic. Our Multi-AGent machine learNing-based approach for Energy efficienT dynamIc Consolidation (MAGNETIC) is implemented in a modified version of the CloudSim simulator, and considers the energy and delay overheads associated with host power mode transition and VM migration, and is evaluated using power traces collected from various workloads running in real servers and resource utilization logs from cloud data center infrastructures. Results show how our strategy reduces data center energy consumption by up to 15 percent compared to other works in the state-of-the-art (SoA), guaranteeing the same QoS and reducing the number of VM migrations and host power mode transitions by up to 86 and 90 percent, respectively. Moreover, it shows better scalability than all other approaches, taking less than 0.7 percent time overhead to execute for a data center with 1,500 VMs. Finally, our solution is capable of detecting host performance variability due to failures, automatically migrating VMs from failing hosts and draining them from workload.

This paper presents the research directions pursued by the WiPLASH European project, pioneering on-chip wireless communications as a disruptive enabler towards next-generation computing systems for artificial intelligence (AI). We illustrate the holistic approach driving our research efforts, which encompass expertises and abstraction levels ranging from physical design of embedded graphene antennas to system-level evaluation of wirelessly-communicating heterogeneous systems.

Edge devices must support computationally demanding algorithms, such as neural networks, within tight area/energy budgets. While approximate computing may alleviate these constraints, limiting induced errors remains an open challenge. In this paper, we propose a hardware/software co-design solution via an inexact multiplier, reducing area/power-delay-product requirements by 73/43%, respectively, while still computing exact results when one input is a Fibonacci encoded value. We introduce a retraining strategy to quantize neural network weights to Fibonacci encoded values, ensuring exact computation during inference. We benchmark our strategy on Squeezenet 1.0, DenseNet-121, and ResNet-18, measuring accuracy degradations of only 0.4/1.1/1.7%.

While Deep Learning (DL) is often considered the state-of-the art for Artificial Intel-ligence-based medical decision support, it remains sparsely implemented in clinical practice and poorly trusted by clinicians due to insufficient interpretability of neural network models. We have approached this issue in the context of online detection of epileptic seizures by developing a DL model from EEG signals, and associating certain properties of the model behavior with the expert medical knowledge. This has conditioned the preparation of the input signals, the network architecture, and the post-processing of the output in line with the domain knowledge. Specifically, we focused the discussion on three main aspects: (1) how to aggregate the classification results on signal segments provided by the DL model into a larger time scale, at the seizure-level; (2) what are the relevant frequency patterns learned in the first convolutional layer of different models, and their relation with the delta, theta, alpha, beta and gamma frequency bands on which the visual interpretation of EEG is based; and (3) the identification of the signal waveforms with larger contribution towards the ictal class, according to the activation differences highlighted using the DeepLIFT method. Results show that the kernel size in the first layer determines the interpretability of the extracted features and the sensitivity of the trained models, even though the final performance is very similar after post-processing. Also, we found that amplitude is the main feature leading to an ictal prediction, suggesting that a larger patient population would be required to learn more complex frequency patterns. Still, our methodology was successfully able to generalize patient inter-variability for the majority of the studied population with a classification F1-score of 0.873 and detecting 90% of the seizures.

The increasing adoption of smart systems in our daily life has led to the development of new applications with varying performance and energy constraints, and suitable computing architectures need to be developed for these new applications. In this article, we present gem5-X, a system-level simulation framework, based on gem-5, for architectural exploration of heterogeneous many-core systems. To demonstrate the capabilities of gem5-X, real-time video analytics is used as a case-study. It is composed of two kernels, namely, video encoding and image classification using convolutional neural networks (CNNs). First, we explore through gem5-X the benefits of latest 3D high bandwidth memory (HBM2) in different architectural configurations. Then, using a two-step exploration methodology, we develop a new optimized clustered-heterogeneous architecture with HBM2 in gem5-X for video analytics application. In this proposed clustered-heterogeneous architecture, ARMv8 in-order cluster with in-cache computing engine executes the video encoding kernel, giving 20% performance and 54% energy benefits compared to baseline ARM in-order and Out-of-Order systems, respectively. Furthermore, thanks to gem5-X, we conclude that ARM Out-of-Order clusters with HBM2 are the best choice to run visual recognition using CNNs, as they outperform DDR4-based system by up to 30% both in terms of performance and energy savings.

Next generation workloads, such as genome sequencing, have an astounding impact in the healthcare sector. Sequence alignment, the first step in genome sequencing, has experienced recent breakthroughs, which resulted in next generation sequencing (NGS). As NGS applications are memory bounded with random memory access patterns, we propose the use of high bandwidth memories like 3D stacked HBM2, instead of traditional DRAMs like DDR4, along with energy efficient compute cores to improve both performance and energy efficiency. Three state-of-the-art NGS applications, Bowtie2, BWA-MEM, and HISAT2 are used as case studies to explore and optimize NGS computing architectures. Then, using the gem5-X architectural simulator, we obtain an overall 68 percent performance improvement and 71 percent energy savings using HBM2 instead of DDR4. Furthermore, we propose an architecture based on ARMv8 cores and demonstrate that 16 ARMv8 64-bit OoO cores with HBM2 outperforms 32-cores of Intel Xeon Phi Knights Landing (KNL) processor with 3D stacked memory. Moreover, we show that by using frequency scaling we can achieve up to 59 percent and 61 percent energy savings for ARM in-order and OoO cores, respectively. Lastly, we show that many ARMv8 in-order cores at 1.5GHz match the performance of fewer OoO cores at 2GHz, while attaining 4.5x energy savings.

Modern datacenters need to tackle efficiently the increasing demand for computing resources while minimizing energy usage and monetary costs. Power market operators have recently introduced emerging demand-response programs, in which electricity consumers regulate their power usage following provider requests to reduce monetary costs. Among different programs, regulation service (RS) reserves are particularly promising for datacenters due to the high credit gain possibilities and datacenters' flexibility in regulating their power consumption. Therefore, it is essential to develop bidding strategies for datacenters to participate in emerging power markets together with power management policies that are aware of power market requirements at runtime. In this paper we propose ECOGreen, a holistic strategy to jointly optimize the datacenter RS problem and virtual machine (VM) allocation that satisfies the hour-ahead power market constraints in the presence of electrical energy storage (EES) and renewable energy. We first find the best power and reserve bidding values as well as the number of active servers in a fast analytical way that works well in practice. Then, we present an online adaptive policy that modulates datacenter power consumption by controlling VMs CPU resource limits and efficiently utilizing demand-side EES and renewable power, while guaranteeing quality-of-service (QoS) constraints. Our results demonstrate that ECOGreen can provide 76 percent of the datacenter power consumption on average as reserves to the market, due to largely operating on renewable sources and EES. This translates into ECOGreen saving up to 71 percent electricity costs when compared to other state-of-the-art datacenter electricity cost minimization techniques that participate in the power market.

Dynamic Thermal Management (DTM) has become a major challenge since it directly affects Multiprocessors Systems-on-chip (MPSoCs) performance, power consumption, and reliability. In this work, we propose a transient fan model, enabling adaptive fan speed control simulation for efficient DTM. Our model is validated through a thermal test chip achieving less than 2°C error in the worst case. With multiple fan speeds, however, the DTM design space grows significantly, which can ultimately make conventional solutions impractical. We address this challenge through a reinforcement learning-based solution to proactively determine the number of active cores, operating frequency, and fan speed. The proposed solution is able to reduce fan power by up to 40% compared to a DTM with constant fan speed with less than 1% performance degradation. Also, compared to a state-of-the-art DTM technique our solution improves the performance by up to 19% for the same fan power.

Deeply-scaled three-dimensional (3D) Multi-Processor Systems-on-Chip (MPSoCs) enable high performance and massive communication bandwidth for next-generation computing. However as process nodes shrink, temperature-dependent leakage dramatically increases, and thermal and power management becomes problematic. In this context, Integrated Flow Cell Array (FCA) technology, which consists of inter-tier microfluidic channels, combines on-chip electrochemical power generation and liquid cooling of 3D MPSoCs. When connected to power delivery networks (PDN) of dies, FCAs provide an additional current compensating the voltage drop (IR-drop). In this paper, we evaluate for the first time how the IR-drop reduction and cooling capabilities of FCAs scale with advanced CMOS processes. We develop a frmework to quantify the system-level impact of FCAs at technology nodes from 22nm to 3nm. Our results show that, across all considered nodes, FCAs reduce the peak temperature of a multi-core processor (MCP) and a Machine Learning (ML) accelerator by over 22°C and 35°C, respectively, compared to off-chip direct liquid cooling. Moreover, the low operation voltages and high temperatures at advanced nodes improve up to 2x FCA power generation. Hence, FCAs allow to keep the IR-drop below 5% for both the MCP and ML accelerator, saving over 10% TSV-reserved area, as opposed to using a High-Performance Computing (HPC) MPSoC liquid cooling solution.

Hybrid caches consisting of both SRAM and emerging Non-Volatile Random Access Memory (eNVRAM) bitcells increase cache capacity and reduce power consumption by taking advantage of eNVRAM's small area footprint and low leakage energy. However, they also inherit eNVRAM's drawbacks, including long write latency and limited endurance. To mitigate these drawbacks, many works propose heuristic strategies to allocate memory blocks into SRAM or eNVRAM arrays at runtime based on block content or access pattern. In contrast, this work presents a HW/SW Stack for Hybrid Caches (SHyCache), consisting of a hybrid cache architecture and supporting programming model, reminiscent of those that enable GP-GPU acceleration, in which application variables can be allocated explicitly to the eNVRAM cache, eliminating the need for heuristics and reducing cache access time, power consumption, and area overhead while maintaining maximal cache utilization efficiency and ease of programming. SHyCache improves performance for applications such as neural networks, which contain large numbers of invariant weight values with high read/write access ratios that can be explicitly allocated to the eNVRAM array. We simulate SHyCache on the gem5-X architectural simulator and demonstrate its utility by benchmarking a range of cache hierarchy variations using three neural networks, namely, Inception v4, ResNet-50, and SqueezeNet 1.0. We demonstrate a design space that can be exploited to optimize performance, power consumption, or endurance, depending on the expected use case of the architecture, while demonstrating maximum performance gains of 1.7 /1.4/1.3x and power consumption reductions of 5.1/5.2/5.4x, for Inception/ResNet/SqueezeNet, respectively.

Area and power-constrained edge devices are increasingly utilized to perform compute intensive workloads, necessitating increasingly area and power-efficient accelerators. In this context, in-SRAM computing performs hundreds of parallel operations on spatially local data common in many emerging workloads, while reducing power consumption due to data movement. However, in-SRAM computing faces many challenges, including integration into the existing architecture, arithmetic operation support, data corruption at high operating frequencies, inability to run at low voltages, and low area density. To meet these challenges, this article introduces BLADE, a BitLine Accelerator for Devices on the Edge. BLADE is an in-SRAM computing architecture that utilizes local wordline groups to perform computations at a frequency 2.8× higher than state-of-the-art in-SRAM computing architectures. BLADE is integrated into the cache hierarchy of low-voltage edge devices, and simulated and benchmarked at the transistor, architecture, and software abstraction levels. Experimental results demonstrate performance/energy gains over an equivalent NEON accelerated processor for a variety of edge device workloads, namely, cryptography (4× performance gain/6× energy reduction), video encoding (6×/2×), and convolutional neural networks (3×/1.5×), while maintaining the highest frequency/energy ratio (up to 2.2 Ghz@1V) of any conventional in-SRAM computing architecture, and a low area overhead of less than 8 percent.

The advent of online video streaming applications and services along with the users' demand for high-quality contents require High Efficiency Video Coding (HEVC), which provides higher video quality and more compression at the cost of increased complexity. On one hand, HEVC exposes a set of dynamically tunable parameters to provide trade-offs among Quality-of-Service (QoS), performance, and power consumption of multi-core servers on the video providers' data center. On the other hand, resource management of modern multi-core servers is in charge of adapting system-level parameters, such as operating frequency and multithreading, to deal with concurrent applications and their requirements. Therefore, efficient multi-user HEVC streaming necessitates joint adaptation of application-and system-level parameters. Nonetheless, dealing with such a large and dynamic design space is challenging and difficult to address through conventional resource management strategies. Thus, in this work, we develop a multi-agent Reinforcement Learning framework to jointly adjust application-and system-level parameters at runtime to satisfy the QoS of multi-user HEVC streaming in power-constrained servers. In particular, the design space, composed of all design parameters, is split into smaller independent sub-spaces. Each design sub-space is assigned to a particular agent so that it can explore it faster, yet accurately. The benefits of our approach are revealed in terms of adaptability and quality (with up to to 4× improvements in terms of QoS when compared to a static resource management scheme), and learning time (6× fasterthan an equivalent mono-agent implementation). Finally, we show that the power-capping techniques formulated outperform the hardware-based power capping with respect to quality.

As the computational complexity of applications on the consumer market, such as high-definition video encoding and deep neural networks, become ever more demanding, novel ways to efficiently compute data intensive workloads are being explored. In this context, In-Memory Computing (IMC) solutions, and particularly bitline computing in SRAM, appear promising as they mitigate one of the most energy consuming aspects in computation: data movement. While IMC architectural level characteristics have been defined by the research community, only a few works so far have explored the implementation of such memories at a low level. Furthermore, these proposed solutions are either slow (<1GHz), area hungry (10T SRAM), or suffer from read disturb and corruption issues. Overall, there is no extensive design study considering realistic assumptions at the circuit level. In this work we propose a fast (up to 2.2Ghz), 6T SRAM-based, reliable (no read disturb issues), and wide voltage range (from 0.6 to 1V) IMC architecture using local bitlines. Beyond standard read and write, the proposed architecture can perform copy, addition and shift operations at the array level. As addition is the slowest operation, we propose a modified carry chain adder, providing a 2× carry propagation improvement. The proposed architecture is validated using a 28nm bulk high performances technology PDK with CMOS variability and post-layout simulations. High density SRAM bitcells (0.127μm) enable area efficiency of 59.7% for a 256×128 array, on par with current industrial standards.

Real-time video transcoding has recently raised as a valid alternative to address the ever-increasing demand for video contents in servers' infrastructures in current multi-user environments. High Efficiency Video Coding (HEVC) makes efficient online transcoding feasible as it enhances user experience by providing the adequate video configuration, reduces pressure on the network, and minimizes inefficient and costly video storage. However, the computational complexity of HEVC, together with its myriad of configuration parameters, raises challenges for power management, throughput control, and Quality of Service (QoS) satisfaction. This is particularly challenging in multi-user environments where multiple users with different resolution demands and bandwidth constraints need to be served simultaneously. In this work, we present MAMUT, a multi-agent machine learning approach to tackle these challenges. Our proposal breaks the design space composed of run-time adaptation of the transcoder and system parameters into smaller sub-spaces that can be explored in a reasonable time by individual agents. While working cooperatively, each agent is in charge of learning and applying the optimal values for internal HEVC and system-wide parameters. In particular, MAMUT dynamically tunes Quantization Parameter, selects number of threads per video, and sets the operating frequency with throughput and video quality objectives under compression and power consumption constraints. We implement MAMUT on an enterprise multicore server and compare equivalent scenarios to state-of-the-art alternative approaches. The obtained results reveal that MAMUT consistently attains up to 8× improvement in terms of FPS violations (and thus Quality of Service), 24% power reduction, as well as faster and more accurate adaptation both to the video contents and available resources.

The power density and, consequently, power hungriness of server processors is growing by the day. Traditional air cooling systems fail to cope with such high heat densities, whereas single-phase liquid-cooling still requires high mass flow-rate, high pumping power, and large facility size. On the contrary, in a micro-scale gravity-driven thermosyphon attached on top of a processor, the refrigerant, absorbing the heat, turns into a two-phase mixture. The vapor-liquid mixture exchanges heat with a coolant at the condenser side, turns back to liquid state, and descends thanks to gravity, eliminating the need for pumping power. However, similar to other cooling technologies, thermosyphon efficiency can considerably vary with respect to workload performance requirements and thermal profile, in addition to the platform features, such as packaging and die floorplan. In this work, we first address the workload- and platform-aware design of a two-phase thermosyphon. Then, we propose a thermal-aware workload mapping strategy considering the potential and limitations of a two-phase thermosyphon to further minimize hot spots and spatial thermal gradients. Our experiments, performed on an 8-core Intel Xeon E5 CPU reveal, on average, up to 10°C reduction in thermal hot spots, and 45% reduction in the maximum spatial thermal gradient on the die. Moreover, our design and mapping strategy are able to decrease the chiller cooling power at least by 45%.

RECIPE (REliable power and time-ConstraInts-aware Predictive management of heterogeneous Exascale systems) is a recently started project funded within the H2020 FETHPC programme, which is expressly targeted at exploring new High-Performance Computing (HPC) technologies. RECIPE aims at introducing a hierarchical runtime resource management infrastructure to optimize energy efficiency and minimize the occurrence of thermal hotspots, while enforcing the time constraints imposed by the applications and ensuring reliability for both time-critical and throughput-oriented computation that run on deeply heterogeneous accelerator-based systems. This paper presents a detailed overview of RECIPE, identifying the fundamental challenges as well as the key innovations addressed by the project, which span run-time management, heterogeneous computing architectures, HPC memory/interconnection infrastructures, thermal modelling, reliability, programming models, and timing analysis. For each of these areas, the paper describes the relevant state of the art as well as the specific actions that the project will take to effectively address the identified technological challenges.

The increasing ubiquity of edge devices in the consumer market, along with their ever more computationally expensive workloads, necessitate corresponding increases in computing power to support such workloads. In-memory computing is attractive in edge devices as it reuses preexisting memory elements, thus limiting area overhead. Additionally, in-SRAM Computing (iSC) efficiently performs computations on spatially local data found in a variety of emerging edge device workloads. We therefore propose, implement, and benchmark BLADE, a BitLine Accelerator for Devices on the Edge. BLADE is an iSC architecture that can perform massive SIMD-like complex operations on hundreds to thousands of operands simultaneously. We implement BLADE in 28nm CMOS and demonstrate its functionality down to 0.6V, lower than any conventional state-of-the-art iSC architecture. We also benchmark BLADE in conjunction with a full Linux software stack in the gem5 architectural simulator, providing a robust demonstration of its performance gain in comparison to an equivalent embedded processor equipped with a NEON SIMD co-processor. We benchmark BLADE with three emerging edge device workloads, namely cryptography, high efficiency video coding, and convolutional neural networks, and demonstrate 4x, 6x, and 3x performance improvement, respectively, in comparison to a baseline CPU/NEON processor at an equivalent power budget.

Integrated Flow Cell Array (FCA) technology promises to address the power delivery and heat dissipation challenges in three-dimensional Multi-Processor Systems-on-Chips (3D MPSoCs) by providing combined inter-tier liquid cooling and power generation capabilities. In this paper, we present for the first time a design framework to accurately model the temperature-aware power delivery network in 3D MPSoCs, and quantify the effects of FCAs on the voltage drop (IR-drop). This framework estimates the power generation variation along FCAs due to voltage and temperature, in the case of uniform and non-uniform powermaps from several real processor traces. Furthermore, we explore different 3D MPSoC configurations to quantify their power delivery requirements. Our results show that FCAs improve the IR-drop with respect to state-of-the-art design methods up to 53% and 30% for dies with a power consumption of 60W and 190W, respectively, while maintaining their peak temperatures below 52°C, and at no additional Through Silicon Via (TSV) area overhead. In addition, as the presence of high power density regions (hotspots) can decrease the FCAs IR-drop reduction by up to 21% with respect to the average value, we present a scalable TSV placement optimization methodology using the proposed framework. This methodology minimizes the IR-drop at hotspots and guarantees an optimal and uniform exploitation of the IR-drop reduction benefits of FCAs.

Accurate workload prediction and throughput estimation are keys in efficient proactive power and performance management of multi-core platforms. Although hardware performance counters available on modern platforms contain important information about the application behavior, employing them efficiently is not straightforward when dealing with time-varying applications even if they have iterative structures. In this work, we propose a machine learning-based framework for workload prediction and throughput estimation using hardware events. Our framework enables throughput estimation over various available system configurations, namely, number of parallel threads and operating frequency. In particular, we first employ workload clustering and classification techniques along with Markov chains to predict the next workload for each available system configuration. Then, the predicted workload is used to estimate the next expected throughput through a machine learning-based regression model. The comparison with state of the art demonstrates that our framework is able to improve Quality of Service (QoS) by 3.4x, while consuming 15% less power thanks to the more accurate throughput estimation.

With the spread of cloud services and Internet of Things concept, there is a popularization of machine learning and artificial intelligence based analytics in our everyday life. However, an efficient deployment of these data-intensive services requires performing computations closer to the edge. In this context, in-cache computing, based on bitline computing, is promising to execute data-intensive algorithms in an energy efficient way by mitigating data movement in the cache hierarchy and exploiting data parallelism. Nevertheless, previous in-cache computing architectures contain serious circuit-level deficiencies (i.e., low bitcell density, data corruption risks, and limited performance), thus report high multiplication latency, which is a key operation for machine learning and deep learning. Moreover, no previous work addresses the issue of way misalignment, strongly constraining data placement not to reduce performance gains. In this work we drastically improve the previously proposed BLADE architecture for in-cache computing to efficiently support multiplication operations by enhancing the local bitline circuitry, enabling associativity-agnostic operations as well as in-place shifting inside local bitline groups. We implemented and simulated the proposed architecture in CMOS 28nm bulk technology from TSMC, validating its functionality and extracting its performance, area, and energy per operation. Then, we designed a behavioral model of the proposed architecture to assess its performance with respect to the latest BLADE architecture. We show a 17.5 and 22% area and energy reduction thanks to the proposed LG optimization. Finally, for 16bits multiplication, we demonstrate 44% cycle count, 47% energy and 41% performances gain versus BLADE and show that 4 embedded shifts is the best trade-off between energy, area and performances.

In this work we present ContainEnergy, a new performance evaluation and profiling tool that uses software containers to perform application runtime assessment, providing energy and performance profiling data. It is focused on energy efficiency for next generation workloads and IT infrastructure.

The need for running complex Machine Learning (ML) algorithms, such as Convolutional Neural Networks (CNNs), in edge devices, which are highly constrained in terms of computing power and energy, makes it important to execute such applications efficiently. The situation has led to the popularization of Binary Neural Networks (BNNs), which significantly reduce execution time and memory requirements by representing the weights (and possibly the data being operated) using only one bit. Because approximately 90% of the operations executed by CNNs and BNNs are convolutions, a significant part of the memory transfers consists of fetching the convolutional kernels. Such kernels are usually small (e.g., 3×3 operands), and particularly in BNNs redundancy is expected. Therefore, equal kernels can be mapped to the same memory addresses, requiring significantly less memory to store them. In this context, this paper presents a custom Binary Dot Product Engine (BDPE) for BNNs that exploits the features of Resistive Random-Access Memories (RRAMs). This new engine allows accelerating the execution of the inference phase of BNNs. The novel BDPE locally stores the most used binary weights and performs binary convolution using computing capabilities enabled by the RRAMs. The system-level gem5 architectural simulator was used together with a C-based ML framework to evaluate the system's performance and obtain power results. Results show that this novel BDPE improves performance by 11.3%, energy efficiency by 7.4% and reduces the number of memory accesses by 10.7% at a cost of less than 0.3% additional die area, when integrated with a 28 nm Fully Depleted Silicon On Insulator ARMv8 in-order core, in comparison to a fully-optimized baseline of YoloV3 XNOR-Net running in a unmodified Central Processing Unit.

In order to optimize the energy use of servers in Data Centers, techniques such as power capping or power budgeting are usually deployed. These techniques rely on the prediction of the power and execution time of applications. These data are obtained via dynamic profiling which requires a full execution of the application. This is not feasible in High Performance Computing (HPC) applications with long execution times. In this paper, we present a methodology to estimate the dynamic CPU and memory energy consumption of an application without executing it completely. Our methodology merges static code analysis information and dynamic profiling via the partial execution of the application. We do so by leveraging the concept of application signature, defined as a reduced version of the application in terms of execution time and power profile. We validate our methodology with a set of CPU -intensive, memory-intensive benchmarks and multi-threaded applications in a presently shipping enterprise server. Our energy estimation methodology shows an overall error below 8.0% when compared to the dynamic energy of the whole execution of the application. Also, our energy estimation methodology allows to estimate the energy of multi-threaded applications with an RMSE equal to 12.7% when compared to the dynamic energy from the complete parallel execution.

Bio-medical image processing in the field of telemedicine, and in particular the definition of systems that allow medical diagnostics in a collaborative and distributed way is experiencing an undeniable growth. Due to the high quality of bio-medical videos and the subsequent large volumes of data generated, to enable medical diagnosis on-the-go it is imperative to efficiently transcode and stream the stored videos on real time, without quality loss. However, online video transcoding is a high-demanding computationally-intensive task and its efficient management in Multiprocessor Systems-on-Chip (MPSoCs) poses an important challenge. In this work, we propose an efficient motion- and texture-aware frame-level parallelization approach to enable online medical imaging transcoding on MPSoCs for next generation video encoders. By exploiting the unique characteristics of bio-medical videos and the medical procedure that enable diagnosis, we split frames into tiles based on their motion and texture, deciding the most adequate level of parallelization. Then, we employ the available encoding parameters to satisfy the required video quality and compression. Moreover, we propose a new fast motion search algorithm for bio-medical videos that allows to drastically reduce the computational complexity of the encoder, thus achieving the frame rates required for online transcoding. Finally, we heuristically allocate the threads to the most appropriate available resources and set the operating frequency of each one. We evaluate our work on an enterprise multicore server achieving online medical imaging with 1.6x higher throughput and 44% less power consumption when compared to the state-of-the-art techniques.

Cloud Computing aims to efficiently tackle the increasing demand of computing resources, and its popularity has led to a dramatic increase in the number of computing servers and data centers worldwide. However, as effect of post-Dennard scaling, computing servers have become power-limited, and new system-level approaches must be used to improve their energy efficiency. This paper first presents an accurate power modelling characterization for a new server architecture based on the FD-SOI process technology for near-threshold computing (NTC). Then, we explore the existing energy vs. performance trade-offs when virtualized applications with different CPU utilization and memory footprint characteristics are executed. Finally, based on this analysis, we propose a novel dynamic virtual machine (VM) allocation method that exploits the knowledge of VMs characteristics together with our accurate server power model for next-generation NTC-based data centers, while guaranteeing quality of service (QoS) requirements. Our results demonstrate the inefficiency of current workload consolidation techniques for new NTC-based data center designs, and how our proposed method provides up to 45% energy savings when compared to state-of-the-art consolidation-based approaches.

The design of new streaming systems is becoming a major area of research to deploy services targeted in the Internet-of-Things (IoT) era. In this context, the new High Efficiency Video Coding (HEVC) standard provides high efficiency and scalability of quality at the cost of increased computational complexity for edge nodes, which is a new challenge for the design of IoT systems. The usage of hardware acceleration in conjunction with general-purpose cores in Multiprocessor Systems-on-Chip (MP-SoCs) is a promising solution to create heterogeneous computing systems to manage the complexity of real-time streaming for high-end IoT systems, achieving higher throughput and power efficiency when compared to conventional processors alone. Furthermore, Machine Learning (ML) provides a promising solution to efficiently use this next-generation of heterogeneous MPSoC designs that the EDA industry is developing by dynamically optimizing system performance under diverse requirements such as frame resolution, search area, operating frequency and stream allocation. In this work, we propose an ML-based approach for stream allocation and Dynamic Voltage and Frequency Scaling (DVFS) management on a heterogeneous MPSoC composed of ARM cores and FPGA fabric containing hardware accelerators for the motion estimation of HEVC encoding. Our experiments on a Zynq7000 SoC outline 20% higher throughput when compared to the state-of-the-art streaming systems for next-generation IoT devices.

Integrated flow cell array (FCA) is an emerging technology, targeting the cooling and power delivery challenges of modern 2D/3D Multi-Processor Systems-on-Chip (MPSoCs). In FCA, electrolytic solutions are pumped through microchannels etched in the silicon of the chips, removing heat from the system, while, at the same time, generating power on-chip. In this work, we explore the impact of FCA system design on various 3D architectures and propose a methodology to optimize a 3D MPSoC with integrated FCA to run a given workload in the most energy-efficient way. Our results show that an optimized configuration can save up to 50% energy with respect to sub-optimal 3D MPSoC configurations.

Integrated Flow-Cell Arrays (FCAs) represent a combination of integrated liquid cooling and on-chip power generation, converting chemical energy of the flowing electrolyte solutions to electrical energy. The FCA technology provides a promising way to address both heat removal and power delivery issues in 3D Multiprocessor Systems-on-Chips (MPSoCs). In this paper we motivate the benefits of FCA in 3D MPSoCs via a qualitative analysis and explore the capabilities of the proposed technology using our extended PowerCool simulator. PowerCool is a tool that performs combined compact thermal and electrochemical simulation of 3D MPSoCs with inter-tier FCA-based cooling and power generation. We validate our electrochemical model against experimental data obtained using a micro-scale FCA, and extend PowerCool with a compact thermal model (3D-ICE) and subthreshold leakage estimation. We show the sensitivity of the FCA cooling and power generation on the design-time (FCA geometry) and run-time (fluid inlet temperature, flow rate) parameters. Our results show that we can optimize the FCA to keep maximum chip temperature below 95 °C for an average chip power consumption of 50 W/cm2while generating up to 3.6 W per cm2of chip area.

Modern cloud data centers (DCs) need to tackle efficiently the increasing demand for computing resources and address the energy efficiency challenge. Therefore, it is essential to develop resource provisioning policies that are aware of virtual machine (VM) characteristics, such as CPU utilization and data communication, and applicable in dynamic scenarios. Traditional approaches fall short in terms of flexibility and applicability for large-scale DC scenarios. In this paper, we propose a heuristic- and a machine learning (ML)-based VM allocation method and compare them in terms of energy, quality of service (QoS), network traffic, migrations, and scalability for various DC scenarios. Then, we present a novel hyper-heuristic algorithm that exploits the benefits of both methods by dynamically finding the best algorithm, according to a user-defined metric. For optimality assessment, we formulate an integer linear programming (ILP)-based VM allocation method to minimize energy consumption and data communication, which obtains optimal results, but is impractical at runtime. Our results demonstrate that the ML approach provides up to 24% server-to-server network traffic improvement and reduces execution time by up to $480{\times }$ compared to conventional approaches, for large-scale scenarios. On the contrary, the heuristic outperforms the ML method in terms of energy and network traffic for reduced scenarios. We also show that the heuristic and ML approaches have up to 6% energy consumption overhead compared to ILP-based optimal solution. Our hyper-heuristic integrates the strengths of both the heuristic and the ML methods by selecting the best one during runtime.

The emergence of video streaming applications, together with the users' demand for high-resolution contents, has led to the development of new video coding standards, such as High Efficiency Video Coding (HEVC). HEVC provides high efficiency at the cost of increased complexity. This higher computational burden results in increased power consumption in current multicore servers. To tackle this challenge, algorithmic optimizations need to be accompanied by content-aware application-level strategies, able to reduce power while meeting compression and quality requirements. In this paper, we propose a machine learning-based power and thermal management approach that dynamically learns and selects the best encoding configuration and operating frequency for each of the videos running on multicore servers, by using information from frame compression, quality, encoding time, power, and temperature. In addition, we present a resolution-aware video assignment and migration strategy that reduces the peak and average temperature of the chip while maintaining the desirable encoding time. We implemented our approach in an enterprise multicore server and evaluated it under several common scenarios for video providers. On average, compared to a state-of-the-art technique, for the most realistic scenario, our approach improves BD-PSNR and BD-rate by 0.54 dB, and 8 percent, respectively, and reduces the encoding time, power consumption, and average temperature by 15.3, 13, and 10 percent, respectively. Moreover, our proposed approach enhances BDPSNR and BD-rate compared to the HEVC Test Model (HM), by 1.19 dB and 24 percent, respectively, without any encoding time degradation, when power and temperature constraints are relaxed.

The Horizon 2020 MANGO project aims at exploring deeply heterogeneous accelerators for use in High-Performance Computing systems running multiple applications with different Quality of Service (QoS) levels. The main goal of the project is to exploit customization to adapt computing resources to reach the desired QoS. For this purpose, it explores different but interrelated mechanisms across the architecture and system software. In particular, in this paper we focus on the runtime resource management, the thermal management, and support provided for parallel programming, as well as introducing three applications on which the project foreground will be validated.

The increasing success of Cloud Computing applications and online services has contributed to the unsustainability of data center facilities in terms of energy consumption. Higher resource demand has increased the electricity required by computation and cooling resources, leading to power shortages and outages, specially in urban infrastructures. Current energy reduction strategies for Cloud facilities usually disregard the data center topology, the contribution of cooling consumption and the scalability of optimization strategies. Our work tackles the energy challenge by proposing a temperature-aware VM allocation policy based on a Trust-and-Reputation System (TRS). A TRS meets the requirements for inherently distributed environments such as data centers, and allows the implementation of autonomous and scalable VM allocation techniques. For this purpose, we model the relationships between the different computational entities, synthesizing this information in one single metric. This metric, called reputation, would be used to optimize the allocation of VMs in order to reduce energy consumption. We validate our approach with a state-of-the-art Cloud simulator using real Cloud traces. Our results show considerable reduction in energy consumption, reaching up to 46.16% savings in computing power and 17.38% savings in cooling, without QoS degradation while keeping servers below thermal redlining. Moreover, our results show the limitations of the PUE ratio as a metric for energy efficiency. To the best of our knowledge, this paper is the first approach in combining Trust-and-Reputation systems with Cloud Computing VM allocation.

The increasing demand for computational resources has led to a significant growth of data center facilities. A major concern has appeared regarding energy efficiency and consumption in servers and data centers. The use of flexible and scalable server power models is a must in order to enable proactive energy optimization strategies. This paper proposes the use of Evolutionary Computation to obtain a model for server dynamic power consumption. To accomplish this, we collect a significant number of server performance counters for a wide range of sequential and parallel applications, and obtain a model via Genetic Programming techniques. Our methodology enables the unsupervised generation of models for arbitrary server architectures, in a way that is robust to the type of application being executed in the server. With our generated models, we are able to predict the overall server power consumption for arbitrary workloads, outperforming previous approaches in the state-of-the-art.

The computational and cooling power demands of enterprise servers are increasing at an unsustainable rate. Understanding the relationship between computational power, temperature, leakage, and cooling power is crucial to enable energy-efficient operation at the server and data center levels. This paper develops empirical models to estimate the contributions of static and dynamic power consumption in enterprise servers for a wide range of workloads, and analyzes the interactions between temperature, leakage, and cooling power for various workload allocation policies. We propose a cooling management policy that minimizes the server energy consumption by setting the optimum fan speed during runtime. Our experimental results on a presently shipping enterprise server demonstrate that including leakage awareness in workload and cooling management provides additional energy savings without any impact on performance.

Reducing the energy consumption for computation and cooling in servers is a major challenge considering the data center energy costs today. To ensure energy-efficient operation of servers in data centers, the relationship among computational power, temperature, leakage, and cooling power needs to be analyzed. By means of an innovative setup that enables monitoring and controlling the computing and cooling power consumption separately on a commercial enterprise server, this paper studies temperature-leakage-energy tradeoffs, obtaining an empirical model for the leakage component. Using this model, we design a controller that continuously seeks and settles at the optimal fan speed to minimize the energy consumption for a given workload. We run a customized dynamic load-synthesis tool to stress the system. Our proposed cooling controller achieves up to 9% energy savings and 30W reduction in peak power in comparison to the default cooling control scheme.

Energy consumption in data centers is nowadays a critical objective because of its dramatic environmental and economic impact. Over the last years, several approaches have been proposed to tackle the energy/cost optimization problem, but most of them have failed on providing an analytical model to target both the static and dynamic optimization domains for complex heterogeneous data centers. This paper proposes and solves an optimization problem for the energy-driven configuration of a heterogeneous data center. It also advances in the proposition of a new mechanism for task allocation and distribution of workload. The combination of both approaches outperforms previous published results in the field of energy minimization in heterogeneous data centers and scopes a promising area of research.