Hi ef, To fix the problem, you need to make trapb non-stalling for the O3 machine. Also, if you use simlarge , you should get ipc for blackscholes at about 1IPC if not higher for 4 cores. Note that the beginning of these benchmarks shows bad performance for most of the benchmarks. It takes a few billion instructions for performance to ramp up to a reasonable levle.
-Rick ef wrote: > Hello, > > The real problem as mentioned before by Rick Strong, is that during > simulation too many HW_MFPR and HW_MTPR instructions are being > generated. Any ideas on why so many are produced? Its responsible for > 50% of front engine stall! > > For Blackscholes, they are not many system calls being made, its a > really simple program. Maybe the TLB is to small? Might be a linux > issue? > Any other ideas? > Thanks > > On Fri, Oct 2, 2009 at 10:04 PM, Korey Sewell <[email protected]> wrote: > >> In terms of fetching, >> Note the O3 is fetching a cache line at a time not a instruction at a time. >> So theoretically, you only need to fetch one 64-byte line to "fetch" 8 >> instructions.... >> >> On Fri, Oct 2, 2009 at 10:24 PM, ef <[email protected]> wrote: >> >>> Hello, >>> >>> I am testing Blackscholes 2.1 on M5. Alpha FS Mode. It seems that the >>> IPC is pretty low around .73, I noticed that the fetching rate is >>> pretty low , avg fetch of .98 per cycle, out of a 4 wide machine. >>> Going through traces, everything seems normal except: I get messages >>> like this doing certain spots, which I believe is the culprit: >>> >>> 8039402330674: system.detail_cpu1.fetch: Running stage. >>> 8039402330674: system.detail_cpu1.fetch: There are no more threads >>> available to fetch from. >>> >>> This happens for several cycles, then fetching appears normal. >>> >>> Right now I have two cores running two threads. Is there anyways I can >>> find out more about whats going on? >>> >>> Thanks, >>> EF >>> _______________________________________________ >>> m5-users mailing list >>> [email protected] >>> http://m5sim.org/cgi-bin/mailman/listinfo/m5-users >>> >> >> -- >> - Korey >> >> _______________________________________________ >> m5-users mailing list >> [email protected] >> http://m5sim.org/cgi-bin/mailman/listinfo/m5-users >> >> > _______________________________________________ > m5-users mailing list > [email protected] > http://m5sim.org/cgi-bin/mailman/listinfo/m5-users > > _______________________________________________ m5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
