Search

Accurate performance estimation of future many-node machines is challenging because it requires detailed simulation models of both node and network. However, simulating the full system in detail is unfeasible in terms of compute and memory resources. State-of-the-art techniques use a two-phase approach that combines detailed simulation of a single node with network-only simulation of the full system. We show that these techniques, where the detailed node simulation is done in isolation, are inaccurate because they ignore two important node-level effects: compute time variability, and inter-node communication. We propose a novel three-stage simulation method to allow scalable and accurate many-node simulation, combining native profiling, detailed node simulation and high-level network simulation. By including timing variability and the impact of external nodes, our method leads to more accurate estimates. We validate our technique against measurements on a multi-node cluster, and report an average 6.7% error on 64 nodes (maximum error of 12%), compared to on average 27% error and up to 54% when timing variability and the scaling overhead are ignored. At higher node counts, the prediction error of ignoring variable timings and scaling overhead continues to increase compared to our technique, and may lead to selecting the wrong optimal cluster configuration. Using our technique, we are able to accurately project performance to thousands of nodes within a day of simulation time, using only a single or a few simulation hosts. Our method can be used to quickly explore large many-node design spaces, including node micro-architecture, node count and network configuration.

High performance large scale graph analytics is essential to timely analyze relationships in big data sets. Conventional processor architectures suffer from inefficient resource usage and bad scaling on graph workloads. To enable efficient and scalable graph analysis, Intel developed the Programmable Integrated Unified Memory Architecture (PIUMA). PIUMA consists of many multi-threaded cores, fine-grained memory and network accesses, a globally shared address space and powerful offload engines. This paper presents the PIUMA architecture, and provides initial performance estimations, projecting that a PIUMA node will outperform a conventional compute node by one to two orders of magnitude. Furthermore, PIUMA continues to scale across multiple nodes, which is a challenge in conventional multinode setups.

Power management through dynamic core, cache and frequency adaptation is becoming a necessity in today’s power-constrained many-core environments. Unfortunately, as core count grows, the complexity of both the adaptation hardware and the power management algorithms increases exponentially. This calls for hierarchical solutions, such as on-chip voltage regulators per-tile rather than per-core, along with multi-level power management. As power-driven adaptation of shared resources affects multiple threads at once, the efficiency in a tile-organized many-core processor architecture hinges on the ability to co-schedule compatible threads to tiles in tandem with hardware adaptations per tile and per core. In this paper, we propose a two-tier hierarchical power management methodology to exploit per-tile voltage regulators and clustered last-level caches. In addition, we include a novel thread migration layer that (i) analyzes threads running on the tiled many-core processor for shared resource sensitivity in tandem with core, cache and frequency adaptation, and (ii) co-schedules threads per tile with compatible behavior. On a 256-core setup with 4 cores per tile, we show that adding sensitivity-based thread migration to a two-tier power manager improves system performance by 10% on average (and up to 20%) while using 4× less on-chip voltage regulators. It also achieves a performance advantage of 4.2% on average (and up to 12%) over existing solutions that do not take DVFS sensitivity into account., Peer Reviewed, Postprint (author's final draft)

Power management through dynamic core, cache and frequency adaptation is becoming a necessity in today's power-constrained many-core environments. Unfortunately, as core count grows, the complexity of both the adaptation hardware and the power management algorithms increases. In this paper, we propose a two-tier hierarchical power management methodology to exploit per-tile voltage regulators and clustered last-level caches. In addition, we include a novel thread migration layer that (i) analyzes threads running on the tiled many-core processor for shared resource sensitivity in tandem with core, cache and frequency adaptation, and (ii) co-schedules threads per tile with compatible behavior., Peer Reviewed, Postprint (author's final draft)

Driven by the motivation to expose instruction-level parallelism (ILP), microprocessor cores have evolved from simple, in-order pipelines into complex, superscalar out-of-order designs. By extracting ILP, these processors also enable parallel cache and memory operations as a useful side-effect. Today, however, the growing off-chip memory wall and complex cache hierarchies of many-core processors make cache and memory accesses ever more costly. This increases the importance of extracting memory hierarchy parallelism (MHP), while reducing the net impact of more general, yet complex and power-hungry ILP-extraction techniques. In addition, for multi-core processors operating in power- and energy-constrained environments, energy-efficiency has largely replaced single-thread performance as the primary concern. Based on this observation, we propose a core microarchitecture that is aimed squarely at generating parallel accesses to the memory hierarchy while maximizing energy efficiency. The Load Slice Core extends the efficient in-order, stall-on-use core with a second in-order pipeline that enables memory accesses and address-generating instructions to bypass stalled instructions in the main pipeline. Backward program slices containing address-generating instructions leading up to loads and stores are extracted automatically by the hardware, using a novel iterative algorithm that requires no software support or recompilation. On average, the Load Slice Core improves performance over a baseline in-order processor by 53% with overheads of only 15% in area and 22% in power, leading to an increase in energy efficiency (MIPS/Watt) over in-order and out-of-order designs by 43% and over 4.7x, respectively. In addition, for a power- and area-constrained many-core design, the Load Slice Core outperforms both in-order and out-of-order designs, achieving a 53% and 95% higher performance, respectively, thus providing an alternative direction for future many-core processors.

Modern microprocessors are increasingly power-constrained as a result of slowed supply voltage scaling (end of Dennard scaling) in conjunction with the transistor density scaling (Moore's Law). Existing many-core power management techniques such as chip-wide/per-core DVFS, and core and cache adaptation are quite effective in isolation at moderate to high power budgets. However, for future many-core chip, the existing techniques do not scale well to large core counts, small time slices and stringent power budgets. We need a new solution that combines different adaptation and reconfiguration techniques. In this paper, we present Chrysso, an integrated, scalable and low-overhead power management framework. Chrysso consists of a three-step process: leveraging simple analytical performance and power models, pruning the search space early using local Pareto front generation, followed by global utility-based power allocation. This ensures scalable and effective dynamic adaptation of many-core processors at short time scales along multiple axes, including core, cache and per-core DVFS adaptations. By integrating multiple power management techniques into a common methodology, Chrysso provides significant performance improvements over isolated mechanisms within a given power budget without power-gating cores. On a 64-core system, Chrysso improves system throughput by 1.6× and 1.9× over core-gating at stringent power envelops for multi-program (SPEC) and multi-threaded (PARSEC) workloads, respectively., Peer Reviewed, Postprint (author's final draft)

Managing the limited resources of power and memory bandwidth while improving performance on multicore hardware is challenging. In particular, more cores demand more memory bandwidth, and multi-threaded applications increasingly stress memory systems, leading to more energy consumption. However, we demonstrate that not all memory traffic is necessary. For modern Java programs, 10 to 60% of DRAM writes are useless, because the data on these lines are dead - the program is guaranteed to never read them again. Furthermore, reading memory only to immediately zero initialize it wastes bandwidth. We propose a software/hardware cooperative solution: the memory manager communicates dead and zero lines with cache scrubbing instructions. We show how scrubbing instructions satisfy MESI cache coherence protocol invariants and demonstrate them in a Java Virtual Machine and multicore simulator. Scrubbing reduces average DRAM traffic by 59%, total DRAM energy by 14%, and dynamic DRAM energy by 57% on a range of configurations. Cooperative software/hardware cache scrubbing reduces memory bandwidth and improves energy efficiency, two critical problems in modern systems.