skip to main content

Title: Coordinated CTA Combination and Bandwidth Partitioning for GPU Concurrent Kernel Execution
Contemporary GPUs support multiple kernels to run concurrently on the same streaming multiprocessors (SMs). Recent studies have demonstrated that such concurrent kernel execution (CKE) improves both resource utilization and computational throughput. Most of the prior works focus on partitioning the GPU resources at the cooperative thread array (CTA) level or the warp scheduler level to improve CKE. However, significant performance slowdown and unfairness are observed when latency-sensitive kernels co-run with bandwidth-intensive ones. The reason is that bandwidth over-subscription from bandwidth-intensive kernels leads to much aggravated memory access latency, which is highly detrimental to latency-sensitive kernels. Even among bandwidth-intensive kernels, more intensive kernels may unfairly consume much higher bandwidth than less-intensive ones. In this article, we first make a case that such problems cannot be sufficiently solved by managing CTA combinations alone and reveal the fundamental reasons. Then, we propose a coordinated approach for CTA combination and bandwidth partitioning. Our approach dynamically detects co-running kernels as latency sensitive or bandwidth intensive. As both the DRAM bandwidth and L2-to-L1 Network-on-Chip (NoC) bandwidth can be the critical resource, our approach partitions both bandwidth resources coordinately along with selecting proper CTA combinations. The key objective is to allocate more CTA resources for latency-sensitive kernels and more NoC/DRAM bandwidth resources to NoC-/DRAM-intensive kernels. We achieve it using a variation of dominant resource fairness (DRF). Compared with two state-of-the-art CKE optimization schemes, SMK [52] and WS [55], our approach improves the average harmonic speedup by 78% and 39%, respectively. Even compared to the best possible CTA combinations, which are obtained from an exhaustive search among all possible CTA combinations, our approach improves the harmonic speedup by up to 51% and 11% on average.  more » « less
Award ID(s):
1717550 1618509
Author(s) / Creator(s):
; ; ;
Date Published:
Journal Name:
ACM Transactions on Architecture and Code Optimization
Page Range / eLocation ID:
1 to 27
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. Die-stacked DRAM (a.k.a., on-chip DRAM) provides much higher bandwidth and lower latency than off-chip DRAM. It is a promising technology to break the “memory wall”. Die-stacked DRAM can be used either as a cache (i.e., DRAM cache) or as a part of memory (PoM). A DRAM cache design would suffer from more page faults than a PoM design as the DRAM cache cannot contribute towards capacity of main memory. At the same time, obtaining high performance requires PoM systems to swap requested data to the die-stacked DRAM. Existing PoM designs fall into two categories - line-based and page-based. The former ensures low off-chip bandwidth utilization but suffers from a low hit ratio of on-chip memory due to limited temporal locality. In contrast, page-based designs achieve a high hit ratio of on-chip memory albeit at the cost of moving large amounts of data between on-chip and off-chip memories, leading to increased off-chip bandwidth utilization and significant system performance degradation. To achieve a similar high hit ratio of on-chip memory as pagebased designs, and eliminate excessive off-chip traffic involved, we propose SELF, a high performance and bandwidth efficient approach. The key idea is to SElectively swap Lines in a requested page that are likely to be accessed according to page Footprint, instead of blindly swapping an entire page. In doing so, SELF allows incoming requests to be serviced from the on-chip memory as much as possible, while avoiding swapping unused lines to reduce memory bandwidth consumption. We evaluate a memory system which consists of 4GB on-chip DRAM and 12GB offchip DRAM. Compared to a baseline system that has the same total capacity of 16GB off-chip DRAM, SELF improves the performance in terms of instructions per cycle by 26.9%, and reduces the energy consumption per memory access by 47.9% on average. In contrast, state-of-the-art line-based and page-based PoM designs can only improve the performance by 9.5% and 9.9%, respectively, against the same baseline system. 
    more » « less
  2. State-of-the-art systems, whether in servers or desktops, provide ample computational and storage resources to allow multiple simultaneously executing potentially parallel applications. However, performance tends to be unpredictable, being a function of algorithmic design, resource allocation choices, and hardware resource limitations. In this article, we introduce MAPPER, a manager of application performance via parallel efficiency regulation. MAPPER uses a privileged daemon to monitor (using hardware performance counters) and coordinate all participating applications by making two coupled decisions: the degree of parallelism to allow each application to improve system efficiency while guaranteeing quality of service (QoS), and which specific CPU cores to schedule applications on. The QoS metric may be chosen by the application and could be in terms of execution time, throughput, or tail latency, relative to the maximum performance achievable on the machine. We demonstrate that using a normalized parallel efficiency metric allows comparison across and cooperation among applications to guarantee their required QoS. While MAPPER may be used without application or runtime modification, use of a simple interface to communicate application-level knowledge improves MAPPER’s efficacy. Using a QoS guarantee of 85% of the IPC achieved with a fair share of resources on the machine, MAPPER achieves up to 3.3 \( \times \) speedup relative to unmodified Linux and runtime systems, with an average improvement of 17% in our test cases. At the same time, MAPPER violates QoS for only 2% of the applications (compared to 23% for Linux), while placing much tighter bounds on the worst case. MAPPER relieves hardware bottlenecks via task-to-CPU placement and allocates more CPU contexts to applications that exhibit higher parallel efficiency while guaranteeing QoS, resulting in both individual application performance predictability and overall system efficiency. 
    more » « less
  3. Network quality-of-service (QoS) does not always translate to user quality-of-experience (QoE). Consequently, knowledge of user QoE is desirable in several scenarios that have traditionally operated on QoS information. Examples include traffic management by ISPs and resource allocation by the operating system. But today these systems lack ways to measure user QoE. To help address this problem, we propose offline generation of per-app models mapping app-independent QoS metrics to app-specific QoE metrics. This enables any entity that can observe an app's network traffic-including ISPs and access points-to infer the app's QoE. We describe how to generate such models for many diverse apps with significantly different QoE metrics. We generate models for common user interactions of 60 popular apps. We then demonstrate the utility of these models by implementing a QoE-aware traffic management framework and evaluate it on a WiFi access point. Our approach successfully improves QoE metrics that reflect user-perceived performance. First, we demonstrate that prioritizing traffic for latency-sensitive apps can improve responsiveness and video frame rate, by 46% and 115%, respectively. Second, we show that a novel QoE-aware bandwidth allocation scheme for bandwidth-intensive apps can improve average video bitrate for multiple users by up to 23%. 
    more » « less
  4. Phase Change Memory (PCM) is an attractive candidate for main memory, as it offers non-volatility and zero leakage power while providing higher cell densities, longer data retention time, and higher capacity scaling compared to DRAM. In PCM, data is stored in the crystalline or amorphous state of the phase change material. The typical electrically controlled PCM (EPCM), however, suffers from longer write latency and higher write energy compared to DRAM and limited multi-level cell (MLC) capacities. These challenges limit the performance of data-intensive applications running on computing systems with EPCMs.

    Recently, researchers demonstrated optically controlled PCM (OPCM) cells with support for 5bits/cellin contrast to 2bits/cellin EPCM. These OPCM cells can be accessed directly with optical signals that are multiplexed in high-bandwidth-density silicon-photonic links. The higher MLC capacity in OPCM and the direct cell access using optical signals enable an increased read/write throughput and lower energy per access than EPCM. However, due to the direct cell access using optical signals, OPCM systems cannot be designed using conventional memory architecture. We need a complete redesign of the memory architecture that is tailored to the properties of OPCM technology.

    This article presents the design of a unified network and main memory system called COSMOS that combines OPCM and silicon-photonic links to achieve high memory throughput. COSMOS is composed of a hierarchical multi-banked OPCM array with novel read and write access protocols. COSMOS uses an Electrical-Optical-Electrical (E-O-E) control unit to map standard DRAM read/write commands (sent in electrical domain) from the memory controller on to optical signals that access the OPCM cells. Our evaluation of a 2.5D-integrated system containing a processor and COSMOS demonstrates2.14 ×average speedup across graph and HPC workloads compared to an EPCM system. COSMOS consumes3.8×lower read energy-per-bit and5.97×lower write energy-per-bit compared to EPCM. COSMOS is the first non-volatile memory that provides comparable performance and energy consumption as DDR5 in addition to increased bit density, higher area efficiency, and improved scalability.

    more » « less
  5. Compute heterogeneity is increasingly gaining prominence in modern datacenters due to the addition of accelerators like GPUs and FPGAs. We observe that datacenter schedulers are agnostic of these emerging accelerators, especially their resource utilization footprints, and thus, not well equipped to dynamically provision them based on the application needs. We observe that the state-of-the-art datacenter schedulers fail to provide fine-grained resource guarantees for latency-sensitive tasks that are GPU-bound. Specifically for GPUs, this results in resource fragmentation and interference leading to poor utilization of allocated GPU resources. Furthermore, GPUs exhibit highly linear energy efficiency with respect to utilization and hence proactive management of these resources is essential to keep the operational costs low while ensuring the end-to-end Quality of Service (QoS) in case of user-facing queries.Towards addressing the GPU orchestration problem, we build Knots, a GPU-aware resource orchestration layer and integrate it with the Kubernetes container orchestrator to build Kube- Knots. Kube-Knots can dynamically harvest spare compute cycles through dynamic container orchestration enabling co-location of latency-critical and batch workloads together while improving the overall resource utilization. We design and evaluate two GPU-based scheduling techniques to schedule datacenter-scale workloads through Kube-Knots on a ten node GPU cluster. Our proposed Correlation Based Prediction (CBP) and Peak Prediction (PP) schemes together improves both average and 99 th percentile cluster-wide GPU utilization by up to 80% in case of HPC workloads. In addition, CBP+PP improves the average job completion times (JCT) of deep learning workloads by up to 36% when compared to state-of-the-art schedulers. This leads to 33% cluster-wide energy savings on an average for three different workloads compared to state-of-the-art GPU-agnostic schedulers. Further, the proposed PP scheduler guarantees the end-to-end QoS for latency-critical queries by reducing QoS violations by up to 53% when compared to state-of-the-art GPU schedulers. 
    more » « less