Search

When multiple processor cores (CPUs) and a GPU integrated together on the same chip share the off-chip DRAM, requests from the GPU can heavily interfere with requests from the CPUs, leading to low system performance and starvation of cores. Unfortunately, state-of-the-art memory scheduling algorithms are ineffective at solving this problem due to the very large amount of GPU memory traffic, unless a very large and costly request buffer is employed to provide these algorithms with enough visibility across the global request stream. Previously-proposed memory controller (MC) designs use a single monolithic structure to perform three main tasks. First, the MC attempts to schedule together requests to the same DRAM row to increase row buffer hit rates. Second, the MC arbitrates among the requesters (CPUs and GPU) to optimize for overall system throughput, average response time, fairness and quality of service. Third, the MC manages the low-level DRAM command scheduling to complete requests while ensuring compliance with all DRAM timing and power constraints. This paper proposes a fundamentally new approach, called the Staged Memory Scheduler (SMS), which decouples the three primary MC tasks into three significantly simpler structures that together improve system performance and fairness. Our evaluation shows that SMS provides 41.2% performance improvement and fairness improvement compared to the best previous state-of-the-art technique, while enabling a design that is significantly less complex and more power-efficient to implement.

In a modern GPU architecture, all threads within a warp execute the same instruction in lockstep. For a memory instruction, this can lead to memory divergence: the memory requests for some threads are serviced early, while the remaining requests incur long latencies. This divergence stalls the warp, as it cannot execute the next instruction until all requests from the current instruction complete. In this work, we make three new observations. First, GPGPU warps exhibit heterogeneous memory divergence behavior at the shared cache: some warps have most of their requests hit in the cache, while other warps see most of their request miss. Second, a warp retains the same divergence behavior for long periods of execution. Third, requests going to the shared cache can incur queuing delays as large as hundreds of cycles, exacerbating the effects of memory divergence. We propose a set of techniques, collectively called Memory Divergence Correction (MeDiC), that reduce the negative performance impact of memory divergence and cache queuing. MeDiC delivers an average speedup of 21.8%, and 20.1% higher energy efficiency, over a state-of-the-art GPU cache management mechanism across 15 different GPGPU applications.

Recent studies have demonstrated that near-data processing (NDP) is an effective technique for improving performance and energy efficiency of data-intensive workloads. However, leveraging NDP in realistic systems with multiple memory modules introduces a new challenge. In today's systems, where no computation occurs in memory modules, the physical address space is interleaved at a fine granularity among all memory modules to help improve the utilization of processor-memory interfaces by distributing the memory traffic. However, this is at odds with efficient use of NDP, which requires careful placement of data in memory modules such that near-data computations and their exclusively used data can be localized in individual memory modules, while distributing shared data among memory modules to reduce hotspots. In order to address this new challenge, we propose a set of techniques that (1) enable collections of OS pages to either be fine-grain interleaved among memory modules (as is done today) or to be placed contiguously on individual memory modules (as is desirable for NDP private data), and (2) decide whether to localize or distribute each memory object based on its anticipated access pattern and steer computations to the memory where the data they access is located. Our evaluations across a wide range of workloads show that the proposed mechanism improves performance by 31% and reduces 38% remote data accesses over a baseline system that cannot exploit computate-data affinity characteristics., Comment: 14 pages, 16 figures

Hierarchical ring networks, which hierarchically connect multiple levels of rings, have been proposed in the past to improve the scalability of ring interconnects, but past hierarchical ring designs sacrifice some of the key benefits of rings by introducing more complex in-ring buffering and buffered flow control. Our goal in this paper is to design a new hierarchical ring interconnect that can maintain most of the simplicity of traditional ring designs (no in-ring buffering or buffered flow control) while achieving high scalability as more complex buffered hierarchical ring designs. Our design, called HiRD (Hierarchical Rings with Deflection), includes features that allow us to mostly maintain the simplicity of traditional simple ring topologies while providing higher energy efficiency and scalability. First, HiRD does not have any buffering or buffered flow control within individual rings, and requires only a small amount of buffering between the ring hierarchy levels. When inter-ring buffers are full, our design simply deflects flits so that they circle the ring and try again, which eliminates the need for in-ring buffering. Second, we introduce two simple mechanisms that provides an end-to-end delivery guarantee within the entire network without impacting the critical path or latency of the vast majority of network traffic. HiRD attains equal or better performance at better energy efficiency than multiple versions of both a previous hierarchical ring design and a traditional single ring design. We also analyze our design's characteristics and injection and delivery guarantees. We conclude that HiRD can be a compelling design point that allows higher energy efficiency and scalability while retaining the simplicity and appeal of conventional ring-based designs.

Die-stacked DRAM has been proposed for use as a large, high-bandwidth, last-level cache with hundreds or thousands of megabytes of capacity. Not all workloads (or phases) can productively utilize this much cache space, however. Unfortunately, the unused (or under-used) cache continues to consume power due to leakage in the peripheral circuitry and periodic DRAM refresh. Dynamically adjusting the available DRAM cache capacity could largely eliminate this energy overhead. However, the current proposed DRAM cache organization introduces new challenges for dynamic cache resizing. The organization differs from a conventional SRAM cache organization because it places entire cache sets and their tags within a single bank to reduce on-chip area and power overhead. Hence, resizing a DRAM cache requires remapping sets from the powered-down banks to active banks. In this paper, we propose CRUNCH (Cache Resizing Using Native Consistent Hashing), a hardware data remapping scheme inspired by consistent hashing, an algorithm originally proposed to uniformly and dynamically distribute Internet traffic across a changing population of web servers. CRUNCH provides a load-balanced remapping of data from the powered-down banks alone to the active banks, without requiring sets from all banks to be remapped, unlike naive schemes to achieve load balancing. CRUNCH remaps only sets from the powered-down banks, so it achieves this load balancing with low bank power-up/down transition latencies. CRUNCH's combination of good load balancing and low transition latencies provides a substrate to enable efficient DRAM cache resizing., Comment: 13 pages