This is a great presentation that goes over modern hardware. It's primarily about cache misses and their impact on performance. Below are some notes on the presentation (time - note).
14:30 - cache hit take 2/3 clocks - miss to memory take 200/300 clocks - 100X cost
15:20 - in multicore you hit l3 because of bandwidth & 1 ft of wire is 1 ghz clock
18 minutes - shadow processing; kind of how the cray does ii
25:30 - out of order execution & cache miss
30 - results - 7 ops out of 300 due to cache miss
33 - miss rates are low; but a tiny (5%) missrate dominates performance
52:20 - cahce misses are hard to detect; they just look like busy cpu top doesn't help...
2 comments
[ 2.7 ms ] story [ 11.6 ms ] thread14:30 - cache hit take 2/3 clocks - miss to memory take 200/300 clocks - 100X cost
15:20 - in multicore you hit l3 because of bandwidth & 1 ft of wire is 1 ghz clock
18 minutes - shadow processing; kind of how the cray does ii
25:30 - out of order execution & cache miss
30 - results - 7 ops out of 300 due to cache miss
33 - miss rates are low; but a tiny (5%) missrate dominates performance
52:20 - cahce misses are hard to detect; they just look like busy cpu top doesn't help...