I just want to confirm that this is true:

In increasing the cache size, so that it starts to cover more of the a working set but not all of it (so if you were moving between the drop off after the first working set to right before the drop off for the second working set), you can store more things in your cache, so the data traffic decreases because of fewer capacity misses. However, it does not decrease as much because since you are not covering the entire working set, there will be conflict misses that replace cache lines. Is this the reason why it does not increase as dramatically as it does after covering the entire working set?

This comment was marked helpful 0 times.

@ycp: yes, that's a very plausible description of behavior that could create this graph.

Another situation would be a case where the access pattern is streaming over an array over and over again:

while (!done) {
  for (int i=0; i<N; i++) {
     // do something with A[i]...
  }

If the array is larger than the cache, all values of A will need to be fetched from memory (due to capacity misses). If A fits in cache there there are no misses. So in this case instead of trending down slightly the curve would be completely flat until the cache size could hold all of A, then it would drop sharply.

This comment was marked helpful 0 times.