A Tool Environment for Efficient Execution of Shared Memory Programs on NUMA Systems

One of the most important performance issues on NUMA systems is data locality since remote memory accesses have latencies several magnitudes higher than local memory accesses. This paper presents a tool environment targeting at tuning NUMA-based shared memory applications towards better memory locality. This tool environment comprises tools, supporting system facilities, and their interface. Tools include a Data Layout Visualizer (DLV) which offers a program’s memory access histogram in an easy-to-use way allowing an explicite optimization of applications; a Run-time Adaptive System (RAS) which analyzes the low-level memory transactions and redistributes data among processors on-the-fly; and a SIMulation Tool (SIMT), an execution driven multithreads simulator for architecture design and performance evaluation. Experiments on a 4 node NUMA cluster shows that optimization can improve speedup up to 181 factors for small codes. Simulation results show that specific architectures can reduce remote memory accesses as high as 48%.