Multithread Content Based File Chunking System in CPU-GPGPU Heterogeneous Architecture

The fast development of Graphics Processing Unit (GPU) leads to the popularity of General-purpose usage of GPU (GPGPU). So far, most modern computers are CPU-GPGPU heterogeneous architecture and CPU is used as host processor. In this work, we promote a multithread file chunking prototype system, which is able to exploit the hardware organization of the CPU-GPGPU heterogeneous computer and determine which device should be used to chunk the file to accelerate the content based file chunking operation of deduplication. We built rules for the system to choose which device should be used to chunk file and also found the optimal choice of other related parameters of both CPU and GPGPU subsystem like segment size and block dimension. This prototype was implemented and tested. The result of using GTX460(336 cores) and Intel i5 (four cores) shows that this system can increase the chunking speed 63% compared to using GPGPU alone and 80% compared to using CPU alone.

I. Introduction

Accompany with the fast development of computing ability, memory space and I/O bandwidth of massive multi-core Graphics Processing Unit (GPU), the general-purpose usage of GPU (GPGPU) has become popular [1]. The huge number of cores in GPGPU (latest GTX590 contains 1024 cores) makes it efficient in processing large amount of data and providing high parallelism. However, some specific properties of GPGPU like memory hierarchy make it impossible to substitute CPU which leads to the conclusion that the CPU-GPGPU heterogeneous architecture will last for a long time. In a CPU-GPGPU heterogeneous architecture computer, CPU and GPGPU are integrated and CPU is used as the host processor.