[parsec-users] PARSEC on Simics

Jae W. Lee jaewlee at gmail.com
Wed Oct 29 11:23:55 EDT 2008


I think it would be good to post my question in this thread.

I ran streamcluster to completion with "simsmall" input with Ruby
(=the detailed memory hierarchy model in GEMS) turned on. As Javier
pointed out, I also observed that most of data cache misses occurred
in Supervisor mode. Is it a normal behavior of the program?

  L1D_cache_access_mode_type_SupervisorMode:   66657467    99.3255%
  L1D_cache_access_mode_type_UserMode:   452661    0.674505%

  L1I_cache_access_mode_type_SupervisorMode:   33596    12.4701%
  L1I_cache_access_mode_type_UserMode:   235816    87.5299%

(These numbers only account for cache misses but not cache hits
because I set REMOVE_SINGLE_CYCLE_DCACHE_FAST_PATH to be false.)

BTW, is there anyone who was able to run a PARSEC benchmark to
completion with both Ruby and Opal (=the detailed processor model in
GEMS) turned on?  I'd like to know a working configuration of
Simics/GEMS to run the PARSEC benchmark suite.

Thanks,
Jae


On Fri, Sep 12, 2008 at 7:34 AM,  <jlira at ac.upc.edu> wrote:
> Many thanks Christian for responding so quickly. This information clearly solves
> the strange behaviour I was facing on my simulations. I will try following your
> instructions.
>
> Best regards,
>
> Javier
>
> Mensaje citado por Christian Bienia <cbienia at CS.Princeton.EDU>:
>
>> Javier,
>>
>> I doubt that 100M instructions are sufficient to get accurate results with
>> most PARSEC workloads. From the list you gave it seems that workloads that
>> have only one kernel (i.e. almost all the time is spent in only one loop)
>> give you accurate results, whereas workloads with multiple kernels don't.
>> Most likely that's because 100M instructions is not sufficient to make it
>> through all kernels. This can have a high impact on the characteristics of
>> the program.
>>
>> To simulate these workloads accurately, you have to switch between accurate
>> simulation and fast-forward mode so that every kernel is executed at some
>> point. You also need to make sure that the kernels are weighted like in the
>> real program.
>>
>> - Chris
>>
>>
>>
>> -----Original Message-----
>> From: parsec-users-bounces at lists.cs.princeton.edu
>> [mailto:parsec-users-bounces at lists.cs.princeton.edu] On Behalf Of
>> jlira at ac.upc.edu
>> Sent: Thursday, September 11, 2008 2:16 PM
>> To: PARSEC Users
>> Subject: [parsec-users] PARSEC on Simics
>>
>>
>>
>> Hi all,
>>
>> I use Simics+GEMS for simulating PARSEC workloads, and I am facing some
>> unrealistic results on some of them. The key thing is that I am unable to
>> run
>> the whole program, so I simulate a part of each of them.
>>
>> I start my simulations on the beginning of the ROI and execute a piece of
>> 100M
>> of instructions. Although some of the workloads give excellent results
>> (blackscholes, canneal, dedup and swaptions), there are other in which the
>> number of instructions executed in user mode are ridiculous (I guess because
>> some kind of synchronization in the program) (bodytrack, facesim,
>> fluidanimate,
>> freqmine, streamcluster and x264). Thus, resulting statistics can not be
>> realistic because no program code have been executed.
>>
>> I suppose that some of you use Simics or Simic+GEMS like me, and may help me
>> sharing your expirience on that simulators executing the "problematic"
>> workloads. I also accept advises about methodology that can help me on this
>> stuff.
>>
>> Many thanks to all in advance for your help,
>>
>> Javier Lira
>>
>> _______________________________________________
>> parsec-users mailing list
>> parsec-users at lists.cs.princeton.edu
>> https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
>>
>> _______________________________________________
>> parsec-users mailing list
>> parsec-users at lists.cs.princeton.edu
>> https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
>>
>
>
> _______________________________________________
> parsec-users mailing list
> parsec-users at lists.cs.princeton.edu
> https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
>


More information about the parsec-users mailing list