Graphics Processing Units (GPUs) exploit large amounts of thread-level parallelism to provide high instruction throughput and to efficiently hide long-latency stalls. The resulting high throughput, along with continued programmability improvements, have made GPUs an essential computational resource in many domains. Applications from different domains can have vastly different compute and memory demands on the GPU. In a large-scale computing environment, to efficiently accommodate such wide-ranging demands without leaving GPU resources underutilized, multiple applications can share a single GPU, akin to how multiple applications execute concurrently on a CPU. Multi-application concurrency requires several support mechanisms in both hardware and software. One such key mechanism is virtual memory, which manages and protects the address space of each application. However, modern GPUs lack the extensive support for multi-application concurrency available in CPUs, and as a result suffer from high performance overheads when shared by multiple applications, as we demonstrate. We perform a detailed analysis of which multi-application concurrency support limitations hurt GPU performance the most. We find that the poor performance is largely a result of the virtual memory mechanisms employed in modern GPUs. In particular, poor address translation performance is a key obstacle to efficient GPU sharing. State-of-the-art address translation mechanisms, which were designed for single-application execution, experience significant inter-application interference when multiple applications spatially share the GPU. This contention leads to frequent misses in the shared translation lookaside buffer (TLB), where a single miss can induce long-latency stalls for hundreds of threads. As a result, the GPU often cannot schedule enough threads to successfully hide the stalls, which diminishes system throughput and becomes a first-order performance concern. Based on our analysis, we propose MASK, a new GPU framework that provides low-overhead virtual memory support for the concurrent execution of multiple applications. MASK consists of three novel address-translation-aware cache and memory management mechanisms that work together to largely reduce the overhead of address translation: (1) a token-based technique to reduce TLB contention, (2) a bypassing mechanism to improve the effectiveness of cached address translations, and (3) an application-aware memory scheduling scheme to reduce the interference between address translation and data requests. Our evaluations show that MASK restores much of the throughput lost to TLB contention. Relative to a state-of-the-art GPU TLB, MASK improves system throughput by 57.8%, improves IPC throughput by 43.4%, and reduces application-level unfairness by 22.4%. MASK's system throughput is within 23.2% of an ideal GPU system with no address translation overhead.
more »
« less
This content will become publicly available on May 31, 2024
Lazy Load Scheduling for Mixed-criticality Applications in Heterogeneous MPSoCs
Newly emerging multiprocessor system-on-a-chip (MPSoC) platforms provide hard processing cores with programmable logic (PL) for high-performance computing applications. In this article, we take a deep look into these commercially available heterogeneous platforms and show how to design mixed-criticality applications such that different processing components can be isolated to avoid contention on the shared resources such as last-level cache and main memory.
Our approach involves software/hardware co-design to achieve isolation between the different criticality domains. At the hardware level, we use a scratchpad memory (SPM) with dedicated interfaces inside the PL to avoid conflicts in the main memory. At the software level, we employ a hypervisor to support cache-coloring such that conflicts at the shared L2 cache can be avoided. In order to move the tasks in/out of the SPM memory, we rely on a DMA engine and propose a new CPU-DMA co-scheduling policy, called Lazy Load, for which we also derive the response time analysis. The results of a case study on image processing demonstrate that the contention on the shared memory subsystem can be avoided when running with our proposed architecture. Moreover, comprehensive schedulability evaluations show that the newly proposed Lazy Load policy outperforms the existing CPU-DMA scheduling approaches and is effective in mitigating the main memory interference in our proposed architecture.
more »
« less
- Award ID(s):
- 2008799
- NSF-PAR ID:
- 10482047
- Publisher / Repository:
- ACM
- Date Published:
- Journal Name:
- ACM Transactions on Embedded Computing Systems
- Volume:
- 22
- Issue:
- 3
- ISSN:
- 1539-9087
- Page Range / eLocation ID:
- 1 to 26
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
In this paper, we present RT-Gang: a novel realtime gang scheduling framework that enforces a one-gang-at-atime policy. We find that, in a multicore platform, co-scheduling multiple parallel real-time tasks would require highly pessimistic worst-case execution time (WCET) and schedulability analysis—even when there are enough cores—due to contention in shared hardware resources such as cache and DRAM controller. In RT-Gang, all threads of a parallel real-time task form a real-time gang and the scheduler globally enforces the one-gangat-a-time scheduling policy to guarantee tight and accurate task WCET. To minimize under-utilization, we integrate a state-of-the-art memory bandwidth throttling framework to allow safe execution of best-effort tasks. Specifically, any idle cores, if exist, are used to schedule best-effort tasks but their maximum memory bandwidth usages are strictly throttled to tightly bound interference to real-time gang tasks. We implement RT-Gang in the Linux kernel and evaluate it on two representative embedded multicore platforms using both synthetic and real-world DNN workloads. The results show that RT-Gang dramatically improves system predictability and the overhead is negligible.more » « less
-
Technological advances in long read sequences have greatly facilitated the development of genomics. However, managing and analyzing the raw genomic data that outpaces Moore's Law requires extremely high computational efficiency. On the one hand, existing software solutions can take hundreds of CPU hours to complete human genome alignment. On the other hand, the recently proposed hardware platforms achieve low processing throughput with significant overhead. In this paper, we propose PARC, an Processing-in-Memory architecture for long read pairwise alignment leveraging emerging resistive CAM (content-addressable memory) to accelerate the bottleneck chaining step in DNA alignment. Chaining takes 2-tuple anchors as inputs and identifies a set of correlated anchors as potential alignment candidates. Unlike traditional main memory which organizes relational data structure in a linear address space, PARC stores tuples in two neighboring crossbar arrays with shared row decoder such that column-wise in-memory computational operations and row-wise memory accesses can be performed in-situ in a symmetric crossbar structure. Compared to both software tools and state-of-the-art accelerators, PARC shows significant improvement in alignment throughput and energy efficiency, thanks to the in-site computation capability and optimized data mapping.more » « less
-
The ever-increasing demand for high-performance in the time-critical embedded domain has pushed the adoption of powerful yet unpredictable heterogeneous Systems-on-a-Chip. The shared memory subsystem, which is known to be a major source of unpredictability, has been extensively studied, and many mitigation techniques have been proposed. Among them, performance-counter-based regulation techniques have seen widespread adoption. However, the problem of combining performance-based regulation with time-domain isolation has not received enough attention. In this article, we discuss our current work-in-progress on SHCReg (Software Hardware Co-design Regulator). First, we assess the limitations and benefits of combined CPU and memory budgeting. Next, we outline a full-stack hardware/software co-design architecture that aims at improving the interplay between CPU and memory isolation for mixed-criticality tasks running on the same core.more » « less
-
We present DeepPicar, a low-cost deep neural network based autonomous car platform. DeepPicar is a small scale replication of a real self-driving car called DAVE-2 by NVIDIA. DAVE-2 uses a deep convolutional neural network (CNN), which takes images from a front-facing camera as input and produces car steering angles as output. DeepPicar uses the same network architecture-9 layers, 27 million connections and 250K parameters-and can drive itself in real-time using a web camera and a Raspberry Pi 3 quad-core platform. Using DeepPicar, we analyze the Pi 3's computing capabilities to support end-to-end deep learning based real-time control of autonomous vehicles. We also systematically compare other contemporary embedded computing platforms using the DeepPicar's CNN-based real-time control workload. We find that all tested platforms, including the Pi 3, are capable of supporting the CNN-based real-time control, from 20 Hz up to 100 Hz, depending on hardware platform. However, we find that shared resource contention remains an important issue that must be considered in applying CNN models on shared memory based embedded computing platforms; we observe up to 11.6X execution time increase in the CNN based control loop due to shared resource contention. To protect the CNN workload, we also evaluate state-of-the-art cache partitioning and memory bandwidth throttling techniques on the Pi 3. We find that cache partitioning is ineffective, while memory bandwidth throttling is an effective solution.more » « less
