CUDA Libraries and MPI+OpenMP+CUDA - Prace Training Portal

Performance Measurement and MetricsVisual Profiler…• reports data by SM (with exception of L2 and DRAM counters)• needs more executions to collect everything (so… it is a bit slow)– REMEMBER: blocks and warps are scheduled at run-time!• to compute derivated statistics automatically you need it (and X)Manually inspection…• Re-code your kernel to obtain 3 versions…– a base kernel which performs the desired overall operation (T);– a memory kernel which has the same device memory access patterns as thebase kernel but no math operations (M);– a math kernel which performs the math operations of the base kernel withoutaccessing global memory (O).• Run and evaluate the gputime (be smart ! CUDA event)February 10, 2012 PRACE Winter School 201230