Hi, For my rough v8mt prototype, one thread and concatenated v8 benchmarks it looks like following: 12 seconds for all: Richards: 1178 DeltaBlue: 1212 Crypto: 789 RayTrace: 1732 EarleyBoyer: 1994 RegExp: 213 Splay: 2073 Score: 1081
for original version: 11 seconds for all: Richards: 1176 DeltaBlue: 1251 Crypto: 788 RayTrace: 1646 EarleyBoyer: 2374 RegExp: 371 Splay: 1498 Score: 1141 Back to investigating such a difference for last benchmarks :) Best, Maxim Mossienko On 17 сен, 23:16, Stephen Clibbery <[email protected]> wrote: > Hi, > > Thanks for the reply. I'm beginning to see just how much work it would > be. A simple global search for static variables just doesn't get close > to the amount of work really involved. There's a lot of data not > explicitly marked static (including, as you say, variables defined in > macros). There's probably weeks if not months of work here. > > I also think it will be hard to accurately estimate the performance > impact of TLS without actually doing all the work, because until all > the modules are converted, we probably won't know exactly which call > sites would need to access the TLS data. A starting point might be a > throwaway prototype that just compares direct static data access to > TLS access. It would be hard to simulate appropriate cache-usage > patterns in such a test though. I'm also not sure exactly how to read > the results: without knowing the potential call-sites in the V8 code, > it would be difficult to say exactly how fast one TLS access would > need to be compared to the existing static access. Then there's the > question of whether all platforms would have similarly fast TLS > access. > > Anyway, this (V8 with this patch) would still be my preferred solution > overall, but it's probably not practical for me to put this amount of > work in at this time :-( > > Best Regards, > Steve --~--~---------~--~----~------------~-------~--~----~ v8-users mailing list [email protected] http://groups.google.com/group/v8-users -~----------~----~----~----~------~----~------~--~---
