Please keep playing with it a bit. It sounds like comparing memory
traces is actually the way to go. It appears as though we have a bug,
so we would really like to track it down and fix it instead of working
around it. Are you using all the patches that Steve Reinhardt sent to
the mailing list since the b5 release? They contain cache fixes that
are very important for MP simulations.
Ali
On Apr 27, 2008, at 3:31 AM, Shoaib Akram wrote:
I nailed the problem down a bit. Two things seems to give this error
for most of benchmarks. L3 Cache latency and membus.headercycles.
the kind of stuff i am doing i need to play with header cycles. For
some values of latency and header cycles benchmarks work fine while
for other i get the unmapped error. I am keeping the l3 cache
latency more than the time membus is occupied in case of data and
request transfers.
---- Original message ----
Date: Sat, 26 Apr 2008 06:55:04 -0700
From: "Steve Reinhardt" <[EMAIL PROTECTED]>
Subject: Re: [m5-users] Same Error with Detailed Model
To: "M5 users mailing list" <[email protected]>
On Fri, Apr 25, 2008 at 11:48 PM, Shoaib Akram <[EMAIL PROTECTED]>
wrote:
Attached are some files. No parameter of O3CPU is touched. Some
cache parameters like mshrs are changed. I attach my script file,
the result of tracediff with timing and detailed,call stack and
final part of instruction traces from both timing and detailed
model. Looks like detailed is executing instructions tice and at
different ticks but if we ignore these two factores, the results
of tracediff are very similar.
You missed an important detail on the tracediff command line: you
want
to run "Exec,-ExecTicks" (note the hyphen/minus sign in there) not
"Exec,ExecTicks". By default Exec by itself prints the tick numbers
(as you see) but if you add "-ExecTicks" it suppresses that so that
tracediff can actually pinpoint where the execution diverges.
As Gabe says, the tricky part once you've done that is figuring out
"when" things go wrong on the detailed side, but you do have the line
number of the diff from tracediff and can then rerun just the
detailed
version with ticks enabled to see what tick you're at at the
relevenat
line of the trace. I use combination of "head" and "tail" for that,
though as Gabe says a simple perl script works too.
Steve
_______________________________________________
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
