skip to main content

Title: MASK: Redesigning the GPU Memory Hierarchy to Support Multi-Application Concurrency
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 more » 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. « less
; ; ; ; ; ; ;
Award ID(s):
Publication Date:
Journal Name:
Proceedings of the Twenty-Third International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS '18)
Page Range or eLocation-ID:
503 to 518
Sponsoring Org:
National Science Foundation
More Like this
  1. Concurrent kernel execution on GPU has proven an effective technique to improve system throughput by maximizing the resource utilization. In order to increase programmability and meet the increasing memory requirements of data-intensive applications, current GPUs support Unified Virtual Memory (UVM), which provides a virtual memory abstraction with demand paging. By allowing applications to oversubscribe GPU memory, UVM provides increased opportunities to share GPU resources across applications. However, in the presence of applications with competing memory requirements, GPU sharing can lead to performance degradation due to thrashing. NVIDIA's Multiple Process Service (MPS) offers the capability to space share bare metal GPUs,more »thereby enabling cluster workload managers, such as Slurm, to share a single GPU across MPI ranks with limited control over resource partitioning. However, it is not possible to preempt, schedule, or throttle a running GPU process through MPS. These features would enable new OS-managed scheduling policies to be implemented for GPU kernels to dynamically handle resource contention and offer consistent performance. The contribution of this paper is two-fold. We first show how memory oversubscription can impact the performance of concurrent GPU applications. Then, we propose three methods to transparently mitigate memory interference through kernel preemption and scheduling policies. To implement our policies, we develop our own runtime system (PILOT) to serve as an alternative to NVIDIA's MPS. In the presence of memory over-subscription, we noticed a dramatic improvement in the overall throughput when using our scheduling policies and runtime hints.« less
  2. Graphics processing units (GPUs) are becoming default accelerators in many domains such as high-performance computing (HPC), deep learning, and virtual/augmented reality. Recently, GPUs have also shown significant speedups for a variety of security-sensitive applications such as encryptions. These speedups have largely benefited from the high memory bandwidth and compute throughput of GPUs. One of the key features to optimize the memory bandwidth consumption in GPUs is intra-warp memory access coalescing, which merges memory requests originating from different threads of a single warp into as few cache lines as possible. However, this coalescing feature is also shown to make the GPUsmore »prone to the correlation timing attacks as it exposes the relationship between the execution time and the number of coalesced accesses. Consequently, an attacker is able to correctly reveal an AES private key via repeatedly gathering encrypted data and execution time on a GPU. In this work, we propose a series of defense mechanisms to alleviate such timing attacks by carefully trading off performance for improved security. Specifically, we propose to randomize the coalescing logic such that the attacker finds it hard to guess the correct number of coalesced accesses generated. To this end, we propose to randomize: a) the granularity (called as subwarp) at which warp threads are grouped together for coalescing, and b) the threads selected by each subwarp for coalescing. Such randomization techniques result in three mechanisms: fixed-sized subwarp (FSS), random-sized subwarp (RSS), and random-threaded subwarp (RTS). We find that the combination of these security mechanisms offers 24- to 961-times improvement in the security against the correlation timing attacks with 5 to 28% performance degradation. Online copy:« less
  3. Obeid, Iyad ; Selesnick, Ivan ; Picone, Joseph (Ed.)
    The goal of this work was to design a low-cost computing facility that can support the development of an open source digital pathology corpus containing 1M images [1]. A single image from a clinical-grade digital pathology scanner can range in size from hundreds of megabytes to five gigabytes. A 1M image database requires over a petabyte (PB) of disk space. To do meaningful work in this problem space requires a significant allocation of computing resources. The improvements and expansions to our HPC (highperformance computing) cluster, known as Neuronix [2], required to support working with digital pathology fall into two broadmore »categories: computation and storage. To handle the increased computational burden and increase job throughput, we are using Slurm [3] as our scheduler and resource manager. For storage, we have designed and implemented a multi-layer filesystem architecture to distribute a filesystem across multiple machines. These enhancements, which are entirely based on open source software, have extended the capabilities of our cluster and increased its cost-effectiveness. Slurm has numerous features that allow it to generalize to a number of different scenarios. Among the most notable is its support for GPU (graphics processing unit) scheduling. GPUs can offer a tremendous performance increase in machine learning applications [4] and Slurm’s built-in mechanisms for handling them was a key factor in making this choice. Slurm has a general resource (GRES) mechanism that can be used to configure and enable support for resources beyond the ones provided by the traditional HPC scheduler (e.g. memory, wall-clock time), and GPUs are among the GRES types that can be supported by Slurm [5]. In addition to being able to track resources, Slurm does strict enforcement of resource allocation. This becomes very important as the computational demands of the jobs increase, so that they have all the resources they need, and that they don’t take resources from other jobs. It is a common practice among GPU-enabled frameworks to query the CUDA runtime library/drivers and iterate over the list of GPUs, attempting to establish a context on all of them. Slurm is able to affect the hardware discovery process of these jobs, which enables a number of these jobs to run alongside each other, even if the GPUs are in exclusive-process mode. To store large quantities of digital pathology slides, we developed a robust, extensible distributed storage solution. We utilized a number of open source tools to create a single filesystem, which can be mounted by any machine on the network. At the lowest layer of abstraction are the hard drives, which were split into 4 60-disk chassis, using 8TB drives. To support these disks, we have two server units, each equipped with Intel Xeon CPUs and 128GB of RAM. At the filesystem level, we have implemented a multi-layer solution that: (1) connects the disks together into a single filesystem/mountpoint using the ZFS (Zettabyte File System) [6], and (2) connects filesystems on multiple machines together to form a single mountpoint using Gluster [7]. ZFS, initially developed by Sun Microsystems, provides disk-level awareness and a filesystem which takes advantage of that awareness to provide fault tolerance. At the filesystem level, ZFS protects against data corruption and the infamous RAID write-hole bug by implementing a journaling scheme (the ZFS intent log, or ZIL) and copy-on-write functionality. Each machine (1 controller + 2 disk chassis) has its own separate ZFS filesystem. Gluster, essentially a meta-filesystem, takes each of these, and provides the means to connect them together over the network and using distributed (similar to RAID 0 but without striping individual files), and mirrored (similar to RAID 1) configurations [8]. By implementing these improvements, it has been possible to expand the storage and computational power of the Neuronix cluster arbitrarily to support the most computationally-intensive endeavors by scaling horizontally. We have greatly improved the scalability of the cluster while maintaining its excellent price/performance ratio [1].« less
  4. The Message Passing Interface (MPI) has been the dominant message passing solution for scientific computing for decades. MPI point-to-point communications are highly efficient mechanisms for process-to- process communication. However, MPI performance is slowed by concurrency protections in the MPI library when processes utilize multiple threads. MPI’s current thread-level interface imposes these overheads throughout the library when thread safety is needed. While much work has been done to reduce multithreading overheads in MPI, a solution is needed that reduces the number of messages exchanged in a threaded environment. Partitioned communication is included in the MPI 4.0 standard as an alternative thatmore »addresses the challenges of multithreaded communication in MPI today. Partitioned communication reduces overall message volume by creating a buffer-sharing mechanism between threads such that they can indicate when portions of a communication buffer are available to be sent. Separation of the control and data planes in MPI is enabled by allowing persistent initialization and single occurrence message buffer matching from the indication that the data is ready to be sent. This enables the usage commands (destination, size, etc.) can be set up prior to data buffer readiness with readiness triggered by a simple doorbell/counter later. This approach is useful for future development of MPI operations in environments where traditional networking commands can have performance challenges, like accelerators (GPUs, FPGAs). In this paper,we detail the design and implementation of a layered library (built on top of MPI-3.1) and an integrated Open MPI solution that supports the new, MPI-4.0 partitioned communication feature set. The library will enable applications to use currently released MPI implementations and older legacy libraries to provide partitioned communication support while also enabling further exploration of this new communication model in new applications and use cases. We will compare the designs of the library and native Open MPI support, provide performance results and comparisons between the two approaches, and lessons learned from the implementation of partitioned communication in both library and native forms. We find that the native implementation and library have similar performance with a percentage difference under 0.94% in microbenchmarks and performance within 5% for a partitioned communication enabled proxy application.« less
  5. In the past decade, GPUs have become an important resource for compute-intensive, general-purpose GPU applications such as machine learning, big data analysis, and large-scale simulations. In the future, with the explosion of machine learning and big data, application demands will keep increasing, resulting in more data and computation being pushed to GPUs. However, due to the slowing of Moore’s Law and rising manufacturing costs, it is becoming more and more challenging to add compute resources into a single GPU device to improve its throughput. As a result, spreading work across multiple GPUs is popular in data-centric and scientific applications. Formore »example, Facebook uses 8 GPUs per server in their recent machine learning platform. However, research infrastructure has not kept pace with this trend: most GPU hardware simulators, including gem5, only support a single GPU. Thus, it is hard to study interference between GPUs, communication between GPUs, or work scheduling across GPUs. Our research group has been working to address this shortcoming by adding multi-GPU support to gem5. Here, we discuss the changes that were needed, which included updating the emulated driver, GPU components, and coherence protocol.« less