S-L1

Abstract

Implementing a GPU L1 data cache entirely in software to usurp the hardware L1 cache sounds counter-intuitive. However, we show how a software L1 cache can perform significantly better than the hardware L1 cache for data-intensive streaming (i.e., Big-Data) GPGPU applications. Hardware L1 data caches can perform poorly on current GPUs, because the size of the L1 is far too small and its cache line size is too large given the number of threads that typically need to run in parallel. Our paper makes two contributions. First, we experimentally characterize the performance behavior of modern GPU memory hierarchies and in doing so identify a number of bottlenecks. Secondly, we describe the design and implementation of a software L1 cache, S-L1. On ten streaming GPGPU applications, S-L1 performs 1.9 times faster, on average, when compared to using the default hardware L1, and 2.1 times faster, on average, when compared to using no L1 cache.