Influence of InfiniBand FDR on the performance of remote GPU virtualization

Abstract

Abstract-The use of GPUs to accelerate general-purpose scientific and engineering applications is mainstream today, but their adoption in current high-performance computing clusters is impaired primarily by acquisition costs and power consumption. Therefore, the benefits of sharing a reduced number of GPUs among all the nodes of a cluster can be remarkable for many applications. This approach, usually referred to as remote GPU virtualization, aims at reducing the number of GPUs present in a cluster, while increasing their utilization rate.The performance of the interconnection network is key to achieving reasonable performance results by means of remote GPU virtualization. To this end, several networking technologies with throughput comparable to that of PCI Express have appeared recently. In this paper we analyze the influence of InfiniBand FDR on the performance of remote GPU virtualization, comparing its impact on a variety of GPU-accelerated applications with other networking technologies, such as InfiniBand QDR and Gigabit Ethernet. Given the severe limitations of freely available remote GPU virtualization solutions, the rCUDA framework is used as the case study for this analysis. Results show that the new FDR interconnect, featuring higher bandwidth than its predecessors, allows the reduction of the overhead of using GPUs remotely, thus making this approach even more appealing.

Composable architecture for rack scale big data computing

“…So the predominant communication mode between a compute node and composable GPU and FPGA resources is likely through bulk data transfer. It has been shown by [37] that adequate bandwidth such as those offered by RDMA at FDR data rate (56 Gb/s) already demonstrated superior performance than a locally connected GPU.…”

SLURM Support for Remote GPU Virtualization: Implementation and Performance Study

“…Although remote GPU virtualization has demonstrated very low overhead with respect to a configuration with a local GPU [22], due to its novelty, this technology is not yet supported by the job schedulers that are commonly encountered in production clusters (e.g., SLURM [23], PBSPro [24], MOAB [25], TORQUE [26], LSF [27], OAR [28], MAUI [29], LoadLever [30], Condor [31], and Sun Grid Engine [32]). In particular, a common job scheduler in production today only deals with real GPUs so that, when a job requests a number of nodes equipped with one (or more) GPU(s), the scheduler will try to map that job to nodes that actually own the requested number of GPUs, thus impairing the benefits of GPU virtualization.…”

POSTER: Boosting the performance of remote GPU virtualization using InfiniBand connect-IB and PCIe 3.0

“…Remote GPU virtualization techniques can help increase GPU utilization rates, while reducing acquisition and maintenance costs. For these reasons, many different virtualization solutions are available today, such as rCUDA [4], [5], SnuCL [6], GVirtuS [7], DS-CUDA [8], and VOCL [9].…”