I'm not actually using SPEC, so unfortunately I can't provide any such
information. I've done some more tests and the amount of variation with
what I'm doing seems to be low enough now that I can handle it. If there
was an easy way to disable interrupts I would do that just to be safe,
but since there doesn't seem to be I won't bother. Thanks!
-Addison
On 2/27/2012 7:17 PM, avadh patel wrote:
On Thu, Feb 23, 2012 at 1:59 PM, Addison Mayberry
<[email protected] <mailto:[email protected]>> wrote:
I have now updated to the latest patch, and I am happy to report
that the variations have decreased significantly. Thank you very
much for your work there and for pointing it out to me.
The variation is still there, probably within an acceptable range,
but I would still like to try disabling interrupts just to see the
effects as it may be relevant for my work. Is there an easy way to
do this (or could you point me in the right direction)?
I am not sure how to disable interrupts. But as you mentioned you'r
running spec benchmarks then variation must come from different
source. Please can you list benchmarks that still have high
variations. Also can you check if they are crashing in VM (prints
'segfault' or other error) or not.
- Avadh
-Addison
On 02/16/2012 01:24 PM, avadh patel wrote:
On Thu, Feb 16, 2012 at 9:25 AM, Addison Mayberry
<[email protected] <mailto:[email protected]>> wrote:
Greetings,
I am continuing my mission to get repeatable performance out
of microbenchmarks in MARSS, and right now I'm looking at
timer interrupts as a possible culprit. To explain the
situation a little more, I'm running a microbenchmark that
has a checkpoint just before the interesting part. Upon
reloading from the checkpoint, the program drops into PTLsim
and begins crunching away, giving me periodic statistics
dumps. These yield (sometimes significantly) different
results between runs, which I'm trying to fix.
I'm running single_core, so as far as I understand from
reading some of Avadh's previous posts, the most likely
culprits for this problem are interrupts and I/O. Ignoring
I/O for the moment, I'd like to find a way to disable
interrupts during the duration of my microbenchmark. I've
looked into doing it just using options on the guest OS, but
on my setup that would be pretty tricky. Would it be possible
to keep PTLsim from firing those timer interrupts?
I have pushed some bug fixes in latest 'features' branch. We are
testing these changes across different range of benchmarks. You
can pull the changes from github and help us to validate these
changes. Some people have reported that variation issues with
many spec benchmarks are now resolved. Please report back if your
variation issues are resolved.
- Avadh
Also, any thoughts or advice on the general non-determinism
issue would be appreciated. Thanks!
Sincerely,
Addison
_______________________________________________
http://www.marss86.org
Marss86-Devel mailing list
[email protected]
<mailto:[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
