First, there are two performance counters in GPU, which are “ L2 Compression Success Rate” and “L2 Compression Ratio”. Could someone clarify the differences between these two?
Secondly, I profiled cudaCompressibleMemory, and from the print outs it appears that there is compression and performance improvement.
Success Rate is what percentage of the data sent to L2 was successfully compressed. The Compression Ratio is the 1:N data size reduction of the compressed sectors. I.e. if you send 100 sectors to L2 and 25 were compressed, the rate would be 25%. If those 25% were compressed in half, the ratio would be 50%.