Using remote accelerators to improve the performance of mathematical libraries

Abstract

[EN] Virtualization technologies, such as virtual machines, have demonstrated to provide economic savings to data centers because they increase overall resource utilization. However, in addition to virtualize entire computers, the virtualization of individual devices also provides significant benefits. As an example, the remote GPU virtualization technique moves local GPU computations to remote GPU devices. Nevertheless, it is also possible to offload the computationally intensive CPU parts of an application to remote accelerators. In this work we present a first implementation of a new middleware that uses remote accelerators to perform the computations of scientific libraries, which were initially intended to be executed in the local CPU. Moreover, forwarding the computationally intensive parts of these libraries to remote accelerators is done in a transparent way to applications, not having to modify their source code. The first implementation of the new middleware is based on offloading the FFT and BLAS libraries, which are used to present the benefits of the new proposal by carrying out an in-depth performance evaluation. Results demonstrate that the new middleware is feasible. Moreover, scientific libraries such as FFTW may experience a speed-up larger than 25x, despite of having to transfer data back and forth to the remote server owning the accelerator.