Statistical pattern based modeling of GPU memory access streams

Recent research studies have shown that modern GPU performance is often limited by the memory system performance. Optimizing memory hierarchy performance requires GPU designers to draw design insights based on the cache & memory behavior of end-user applications. Unfortunately, it is often difficult to get access to end-user workloads due to the confidential or proprietary nature of the software/data. Furthermore, the efficiency of early design space exploration of cache & memory systems is often limited due to either the slow speed of detailed simulation techniques or limited scope of state-of-the-art cache analytical models. To enable efficient GPU memory system exploration, we present a novel methodology and framework that statistically models the GPU memory access stream locality. The proposed G-MAP (GPU Memory Access Proxy) framework models the regularity in codelocalized memory access patterns of GPGPU applications and the parallelism in GPU's execution model to create miniaturized memory proxies. We evaluate G-MAP using 18 GPGPU benchmarks and show that G-MAP proxies can replicate cache/memory performance of original applications with over 90% accuracy across over 5000 different L1/L2 cache, prefetcher and memory configurations.