Having perf tests would help measure the impact of GPGMM optimizations. Come up with low-level perf test harness that we can use to benchmark allocation performance between GPGMM and other libraries.