The SPEC2006 workloads are not known for having very large working sets. I'd imagine that you simply captured the whole working set for most of the benchmarks with your initial 4MB configuration. I'd try the experiment with a very small LLC (say, 512kb) just to see what the improvement is in this case.
On Wed, Nov 14, 2012 at 5:13 PM, avadh patel <[email protected]> wrote: > > > > On Wed, Nov 14, 2012 at 1:16 PM, kuniors <[email protected]> wrote: > >> ** >> I am attaching the config files. Thanks, >> >> ------------------------------ >> kuniors >> >> *From:* kuniors <[email protected]> >> *Date:* 2012-11-14 12:48 >> *To:* [email protected] >> *Subject:* performance improvement when doubling LLC capacity >> Hello, >> >> I have done the following simulation with the marss codes. >> I have two L2 config files, and the only difference: >> a. 4MB, 8-way associative, >> b. 8MB, 16-way associative. >> >> And I configured the machine with these two cache configuration. >> >> The simulation results of 27 SPEC2006 show that the performance >> improvement is only a little (around 2%, only 3 of them improved a lot >> (~20%) ) when I double the cache capacity. I think there might be something >> wrong for the improvement is too small. Could you please let me know if I >> am doing something wrong? >> >> > Configuration looks ok. Can you check how much LLC cache access is there? > Also how much is cache access rate? You can use periodic stats to find out > those cache access rates. I suspect that if cache access rates are low for > the portion of benchmark that you are simulating then having large caches > might not affect overall performance. > > - Avadh > >> Thanks, >> kuniors >> >> _______________________________________________ >> http://www.marss86.org >> Marss86-Devel mailing list >> [email protected] >> https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel >> >> > > _______________________________________________ > http://www.marss86.org > Marss86-Devel mailing list > [email protected] > https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel > >
_______________________________________________ http://www.marss86.org Marss86-Devel mailing list [email protected] https://www.cs.binghamton.edu/mailman/listinfo/marss86-devel
