Enhanced XMLRPC-based GPU virtualization

Abstract

increasing demand of using GPUs for general purpose computation, especially to accelerate large volume of data processing, highlights the necessity of virtualizing these expensive co-processors in HPC platforms. Current approaches to GPU virtualization are classified into I/O pass-through, API-Remoting, or hybrid. However, one disadvantage of API-Remoting is that it may not be able to achieve near native performance, specifically when it comes to large volume of data processing. Although some approaches like vCUDA implemented a shared memory model (VMRPC) to improve the data transmission latency, still the behavior of XMLRPC during data encapsulation imposes a non-negligible delay. Based on our study, this delay is tied to insufficient data encapsulation in XMLRPC which is addressed in this paper. This paper first clarifies how optimized data encapsulation would affect the overall performance by accelerating data movement from client to server and second investigates its impact on GPU virtualization. To achieve this purpose, a module was designed and developed to automatically encapsulate the source data based on XMLRPC-String and transfer the data to remote GPU using an optimized packet size with minimum number of packets. Our experimental results showed that, optimized data encapsulation in XMLRPC could speed up the data transfer process up to 4.5X, when per-processing is required, and 7X without pre-processing step.