GPUs are widespread across clusters of compute nodes due to their attractive performance for data parallel codes. However, communicating between GPUs across the cluster is cumbersome when compared to CPU networking implementations. A number of recent works have enabled GPUs to more naturally access the network, but suffer from performance problems, require hidden CPU helper threads, or restrict communications to kernel boundaries. In this paper, we propose GPU Triggered Networking, a novel, GPU-centric networking approach which leverages the best of CPUs and GPUs. In this model, CPUs create and stage network messages and GPUs trigger the network interface when data is ready to send. GPU Triggered Networking decouples these two operations, thereby removing the CPU from the critical path. We illustrate how this approach can provide up to 25% speedup compared to standard GPU networking across microbenchmarks, a Jacobi stencil, an important MPI collective operation, and machine-learning workloads.

--

GPUs,NIC hardware,RDMA networks

• Ciências Físicas - Ciências Naturais

Esta publicação é um output do(s) seguinte(s) projeto(s):

• Towards Natural Interaction and Communication