high performance computing on graphics processing units: hgpu.org

As process technology scales, the number of transistors that can fit in a unit area has increased exponentially. Processor throughput, memory storage, and memory throughput have all been increasing at an exponential pace. As such, DRAM has become an ever-tightening bottleneck for applications with irregular memory access patterns. Computer architects in industry sometimes use ad hoc analytical modeling techniques in lieu of cycle-accurate performance simulation to identify critical design points. Moreover, analytical models can provide clear mathematical relationships for how system performance is affected by individual microarchitectural parameters, something that may be difficult to obtain with a detailed performance simulator. Modern DRAM controllers rely on Out-of-Order scheduling policies to increase row access locality and decrease the overheads of timing constraint delays. This paper proposes a hybrid analytical DRAM performance model that uses memory address traces to predict the DRAM efficiency of a DRAM system when using such a memory scheduling policy. To stress our model, we use a massively multithreaded architecture based upon contemporary GPUs to generate our memory address trace. We test our techniques on a set of real CUDA applications and find our hybrid analytical model predicts the DRAM efficiency to within 15.2% absolute error when arithmetically averaged across all applications.