On Mon, August 6, 2012 9:17 am, Andreas Hansson wrote:
> Hi everyone,
>
> A bit more detail on the mysterious pc-simple-timing that passes in zizzer
> but fails on a whole range of other systems. It turns out the zizzer stats
> generated in the test run are wrong(!)
>
> Here's how I found this out:
>
>
> 1)      Added a flag where the stat is incremented (num_mem_refs and
> store_insts in src/cpu/simple/base.cc)
>
> 2)      Dumped on two systems with the stats .value() printed with a
> dedicated flag
>
> 3)      Compared the traces, and they match! This is where my head started
> to spin... :)
>
> 4)      Ran the pc-simple-timing manually on zizzer with the command found
> in the simout and guess what: the stats are the same as on the other
> systems!
>
> 5)      Tried running the regression again with scons -j1 to avoid any odd
> race conditions, but the stats that come out that way are off, suggesting
> the last instruction (or at least memory access) has simply not happened
> yet the moment the stats are dumped. Odd to say the least.
>
> I have massaged the tests/SConscript to not use scons Execute command, but
> instead subprocess.call, and with this change the regressions on zizzer
> also fail.
>
> In conclusion, either scons or gem5 is doing something odd here. Any scons
> ninja out there that has any idea? I tried adding a sleep of 10 s before
> and after, but that does not seem to help. Could it be related to some bug
> in drain?
>
> Perhaps we can bump the scons version on zizzer just in case?
>

I was recently testing a patch related to x86 when I also noticed the
problem and the stats actually matched the ones that you had posted
earlier. I think it is some case similar to x86 cpuid instruction. Steve
had mentioned another instruction for which he saw difference in traces
from two different runs. I tried to figure if there was something wrong,
but was not able to.

I think if there was some problem with scons / dumping of stats, we would
have seen this problem occur with other tests as well.

--
Nilay

_______________________________________________
gem5-dev mailing list
[email protected]
http://m5sim.org/mailman/listinfo/gem5-dev

Reply via email to