Fractional GPUs: Software-based compute and memory bandwidth reservation for GPUs

Abstract

GPUs are increasingly being used in real-time systems, such as autonomous vehicles, due to the vast performance benefits that they offer. As more and more applications use GPUs, more than one application may need to run on the same GPU in parallel. However, real-time systems also require predictable performance from each individual applications which GPUs do not fully support in a multi-tasking environment. Nvidia recently added a new feature in their latest GPU architecture that allows limited resource provisioning. This feature is provided in the form of a closed-source kernel module called the Multi-Process Service (MPS). However, MPS only provides the capability to partition the compute resources of GPU and does not provide any mechanism to avoid inter-application conflicts within the shared memory hierarchy. In our experiments, we find that compute resource partitioning alone is not sufficient for performance isolation. In the worst case, due to interference from co-running GPU tasks, read/write transactions can observe a slowdown of more than 10x. In this paper, we present Fractional GPUs (FGPUs), a software-only mechanism to partition both compute and memory resources of a GPU to allow parallel execution of GPU workloads with performance isolation. As many details of GPU memory hierarchy are not publicly available, we first reverse-engineer the information through various micro-benchmarks. We find that the GPU memory hierarchy is different from that of the CPU, making it well-suited for page coloring. Based on our findings, we were able to partition both the L2 cache and DRAM for multiple Nvidia GPUs. Furthermore, we show that a better strategy exists for partitioning compute resources than the one used by MPS. An FGPU combines both this strategy and memory coloring to provide superior isolation. We compare our FGPU implementation with Nvidia MPS. Compared to MPS, FGPU reduces the average variation in application runtime, in a multi-tenancy environment, from 135% to 9%. To allow multiple applications to use FGPUs seamlessly, we ported Caffe, a popular framework used for machine learning, to use our FGPU API.