On Fri, 2011-09-30 at 12:30 -0700, Geoffrey Brown wrote: > Some more info to save you from reading through the bug: > > tp4m loads a sequence of 21 popular pages, then loads the same sequence of > pages again in the same order, and reports the average of the minimum times > for each page. In practice, the minimum per-page time is very usually the > second access, in part because of cache hits in the Necko memory cache. > > >From a cache perspective, the access pattern does not reflect real-world use > >in that: > - every page that is loaded is guaranteed to be loaded again > - every page has the same access frequency > - pages are loaded in the same order each time > + a cache miss and then a load from the localhost network is not inherently slower than a cache hit and a load from the cache. tp* runs off localhost backends iirc.
When those 2 cases measure so closely a benchmark isn't going to be very useful. Indeed it is going to be dominated by other things (probably layout). IMO - If we want it to be dominated by cache-savings (so that cache-savings can be separated from the other work going on) then some cost has to be added to the cache miss path. I thought this is why nick did necko net. _______________________________________________ dev-tech-network mailing list [email protected] https://lists.mozilla.org/listinfo/dev-tech-network
