skip to main content


Title: UVM Discard: Eliminating Redundant Memory Transfers for Accelerators
An increasing number of applications benefit from heterogeneous hardware accelerators. Such accelerators often require the application to manually manage memory buffers on devices and transfer data between host and device buffers. A programming model that unifies the virtual address space across the host and devices is appealing because it enables automatic memory transfers and simplifies application-level programming. However, the automatic memory transfers can sometimes be redundant, which decreases performance. NVIDIA’s UVM (unified virtual memory) driver provides a unified virtual address space for CPU-GPU programming. This paper identifies redundant memory transfers (RMTs) as a common performance issue with UVM. To address this issue, this paper proposes a data discard directive, and evaluates two implementations of that directive, UvmDiscard and UvmDiscardLazy. This directive exploits application-level knowledge to avoid RMTs. The implementations were integrated with NVIDIA’s open-source UVM driver to demonstrate their usefulness on real-world CUDA UVM applications. For example, the use of the discard directive increases training throughput by 61.2% on a large deep learning application that oversubscribes GPU memory.  more » « less
Award ID(s):
2008857
NSF-PAR ID:
10427139
Author(s) / Creator(s):
; ; ; ; ;
Date Published:
Journal Name:
2022 IEEE International Symposium on Workload Characterization (IISWC)
Page Range / eLocation ID:
27 to 38
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. The recent introduction of Unified Virtual Memory (UVM) in GPUs offers a new programming model that allows GPUs and CPUs to share the same virtual memory space, which shifts the complex memory management from programmers to GPU driver/ hardware and enables kernel execution even when memory is oversubscribed. Meanwhile, UVM may also incur considerable performance overhead due to tracking and data migration along with special handling of page faults and page table walk. As UVM is attracting significant attention from the research community to develop innovative solutions to these problems, in this paper, we propose a comprehensive UVM benchmark suite named UVMBench to facilitate future research on this important topic. The proposed UVMBench consists of 32 representative benchmarks from a wide range of application domains. The suite also features unified programming implementation and diverse memory access patterns across benchmarks, thus allowing thorough evaluation and comparison with current state-of-the-art. A set of experiments have been conducted on real GPUs to verify and analyze the benchmark suite behaviors under various scenarios. 
    more » « less
  2. Unified Virtual Memory (UVM) was recently introduced with CUDA version 8 and the Pascal GPU. The older CUDA programming style is akin to older large-memory UNIX applications which used to directly load and unload memory segments. Newer CUDA programs have started taking advantage of UVM for the same reasons of superior programmability that UNIX applications long ago switched to assuming the presence of virtual memory. Therefore, checkpointing of UVM has become increasing important, especially as NVIDIA CUDA continues to gain wider popularity: 87 of the top 500 supercomputers in the latest listings use NVIDIA GPUs, with a current trend of ten additional NVIDIA-based supercomputers each year. A new scalable checkpointing mechanism, CRUM (Checkpoint-Restart for Unified Memory), is demonstrated for hybrid CUDA/MPI computations across multiple computer nodes. The support for UVM is particularly attractive for programs requiring more memory than resides on the GPU, since the alternative to UVM is for the application to directly copy memory between device and host. Furthermore, CRUM supports a fast, forked checkpointing, which mostly overlaps the CUDA computation with storage of the checkpoint image in stable storage. The runtime overhead of using CRUM is 6% on average, and the time for forked checkpointing is seen to be a factor of up to 40 times less than traditional, synchronous checkpointing. 
    more » « less
  3. Experience shows that on today's high performance systems the utilization of different acceleration cards in conjunction with a high utilization of all other parts of the system is difficult. Future architectures, like exascale clusters, are expected to aggravate this issue as the number of cores are expected to increase and memory hierarchies are expected to become deeper. One big aspect for distributed applications is to guarantee high utilization of all available resources, including local or remote acceleration cards on a cluster while fully using all the available CPU resources and the integration of the GPU work into the overall programming model. For the integration of CUDA code we extended HPX, a general purpose C++ run time system for parallel and distributed applications of any scale, and enabled asynchronous data transfers from and to the GPU device and the asynchronous invocation of CUDA kernels on this data. Both operations are well integrated into the general programming model of HPX which allows to seamlessly overlap any GPU operation with work on the main cores. Any user defined CUDA kernel can be launched on any (local or remote) GPU device available to the distributed application. We present asynchronous implementations for the data transfers and kernel launches for CUDA code as part of a HPX asynchronous execution graph. Using this approach we can combine all remotely and locally available acceleration cards on a cluster to utilize its full performance capabilities. Overhead measurements show, that the integration of the asynchronous operations (data transfer + launches of the kernels) as part of the HPX execution graph imposes no additional computational overhead and significantly eases orchestrating coordinated and concurrent work on the main cores and the used GPU devices. 
    more » « less
  4. 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, 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. 
    more » « less
  5. null (Ed.)
    As specialized hardware accelerators such as GPUs become increasingly popular, developers are looking for ways to target these platforms with high-level APIs. One promising approach is kernel libraries such as PyTorch or cuML, which provide interfaces that mirror CPU-only counterparts such as NumPy or Scikit-Learn. Unfortunately, these libraries are hard to develop and to adopt incrementally: they only support a subset of their CPU equivalents, only work with datasets that fit in device memory, and require developers to reason about data placement and transfers manually. To address these shortcomings, we present a new approach called offload annotations (OAs) that enables heterogeneous GPU computing in existing workloads with few or no code modifications. An annotator annotates the types and functions in a CPU library with equivalent kernel library functions and provides an offloading API to specify how the inputs and outputs of the function can be partitioned into chunks that fit in device memory and transferred between devices. A runtime then maps existing CPU functions to equivalent GPU kernels and schedules execution, data transfers and paging. In data science workloads using CPU libraries such as NumPy and Pandas, OAs enable speedups of up to 1200⇥ and a median speedup of 6.3⇥ by transparently offloading functions to a GPU using existing kernel libraries. In many cases, OAs match the performance of handwritten heterogeneous implementations. Finally, OAs can automatically page data in these workloads to scale to datasets larger than GPU memory, which would need to be done manually with most current GPU libraries. 
    more » « less