2 ms·
Hi Dan -- Great example. I looked briefly at the source, and wasn't sure whether "pointer_chase" was on or off in your graphs. Or maybe it didn't make a diff
by nkurz 13y ago
Hi Dan --
Great example. I looked briefly at the source, and wasn't sure whether "pointer_chase" was on or off in your graphs. Or maybe it didn't make a difference?
Page-aligned accesses like these also make compulsory misses worse, because prefetchers won’t prefetch beyond a page boundary. But if you have enough data that you’re aligning things to page boundaries, you probably can’t do much about that anyway.
To the contrary, I think this is one of the relatively rare cases that explicit prefetching can help you. But maybe this helps only once your sets are too large for L3?
I wrote a little a few months ago on my attempts to speed up a Stream benchmark for Sandy Bridge that might have some overlap with your post: http://software.intel.com/en-us/forums/topic/480004#comment-1763753 http://software.intel.com/en-us/forums/topic/480004#comment-...