Aah, well I really wish to help make it work, the prefetcher makes all
the difference in these workloads, so I can help with some heavy testing...
Yes, I've tested this with the tip as of today from the development repo.
And the app being simulated segfaults, so it would be tough to track
down in simulation unless you run something predictable like memtestx86
or something like that. I can send you the .ini file and binary if that
would help as it tends to mess up very early in the run. Maybe you can
run gdb in the simulated system? : )
I would guess it's something similar to the prefetch data being stale by
the time it gets to the program. Note that I'm not using any L1
prefetcher but have used the same prefetcher for L1 for many simulations
for the past few months without any problems. It's only when moving it
to the L2 that there's trouble.
Joe
On 6/5/2010 2:43 PM, Steve Reinhardt wrote:
Well, I expect it's because the prefetcher is buggy and hasn't been
heavily tested :-).
Are you using the latest development version?
When you say segfaults, do you mean the app running on the simulated
system segfaults, or m5 segfaults? The former is really bad, as it
means there's some bug in the way prefetches are handled in the cache
hierarchy, which could be painful to track down. The latter is still
not good, but could be a simple bug in the prefetcher itself (note
that someone just posted a fix for a memory leak in the stride
prefetcher to the list recently, and I wouldn't be surprised if there
aren't more little bugs like that).
Steve
On Sat, Jun 5, 2010 at 10:55 AM, Joe Gross<joegr...@umd.edu> wrote:
Hello,
I'm trying again to get the hardware L2 prefetcher to work in an attempt to
model the Nehalem architecture more accurately in FS mode. I enabled the
prefetcher in CacheConfig.py as follows:
-system.l2 = L2Cache(size='2MB')
+system.l2 = L2Cache(size='8MB', assoc=16, latency='49ns',
prefetch_policy='ghb', prefetch_degree=2, prefetcher_size=64)
Now when I run various SPEC CPU2006 benchmarks, I tend to get segfaults in
the benchmarks and in the kernel as well. Switching back to the regular
L2Cache gives correct behavior. I am using PhysicalMemory and all other
defaults. I have compiled kernel v2.6.27 and am using the cross compiler
provided on the website. GemsFDTD/test tends to give a segfault, mcf/test
tends to deadlock after some time. Anyone have some idea why this is
happening or how to fix it?
Joe
_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
_______________________________________________
m5-users mailing list
m5-users@m5sim.org
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
