Packet Compression in GPU Architectures

Abstract

Graphical processing unit (GPU) can support multiple operations in parallel by executing it on multiple thread unit known as warp i.e. multiple threads running the same instruction. Each time miss happens at private cache of Streaming Multiprocessor (SM), the request is migrated over the network to shared L2 cache and then later down to Memory Controller (MC) for supplying memory block. The interconnect delay becomes a bottleneck due to a large number of requests from different SM and multiple replies from the MCs. The compression technique can be used to mitigate the performance bottleneck caused by a large volume of data. In this work, I apply various compression algorithms and propose a new compression scheme, Data Segment Matching (DSM). I apply approximation to the floating-point elements to improve compressibility and develop a prediction model to identify number of approximation bits. I focus on compression techniques to resolve this bottleneck. The evaluations using a cycle accurate simulator show that this scheme improves Instructions per Cycle (IPC) by 12% on an average across various benchmarks with compressibility 50% in integer type benchmarks and 35% in floating-point type benchmarks when the proposed scheme is applied to packet compression in the interconnection network.