Characterizing the Execution Dynamics of GPGPU Applications

General Purpose computing on Graphic Processing Unit (GPGPU) is rapidly gaining adoption in scientific and engineering applications. Considerable research efforts have been dedicated to improving the efficiency of GPU microarchitecture. It is thus essential to understand the execution characteristics of GPUs so as to identify clues for microarchitecture innovations. In this work, we performed detailed microarchitecture simulation for GPGPU workloads. A unique contribution of this work is the emphasis on whole program behaviors. A key observation of this work is that the execution characteristics of certain kernels exhibit repeated patterns. Such patterns are the results of GPU´s thread block based execution model, but can only be identified when there are a sufficient number of blocks. We then show that this number of blocks can be used to accelerate microarchitecture simulation. Experimental results show that on average a 7.62X speedup is attained with an error level below 10%.