snipersim parsec run error

[meet...@gmail.com]

Hi All,

I am able to to compile the parsec and splash benchmarks without any issues. But when I am trying to run the parsec benchmarks, I am facing the following issue. My splash runs don't have any problem (I tried couple of splash benchmarks and I can run them succesfully using the test input without any issue).

Command :- $ ./run-sniper -p parsec-blackscholes -i small -n 4 -c gainestown

Error:- 

[PARSEC] Error: Cannot find file 'SNIPER_APP_LD_PRELOAD=$LD_PRELOAD; unset LD_PRELOAD; /home/rp/sniper/sniper/run-sniper -n 4 -m 'localhost' -d '/home/rp/sniper/sniper/benchmarks' -c gainestown --roi --curdir=/home/rp/sniper/sniper/benchmarks  -- '.

I think the error is coming from the parsec/__init.py__ script, when the subprocess communicate is being called in the run function. But I am not sure which file it is not able to find. I have tried manually untarring the inputsimsmall.tar in the sniper/benchmarks/parsec/parsec-2.1/pkgs/apps/blackscholes/inputs/ directory, still the problem persists.

Do I need to take any further steps to be able to run parsec with sniper?

Thanks in advance.

Reena

[Sergio Rivas]

Reena,

I'm using these days Sniper with PARSEC and until now I have not detected any issues. Your launch command seems correct, so make sure you have set GRAPHITE_ROOT environment variable to the Sniper folder and BENCHMARKS_ROOT to the benchmarks root folder.

Other than that, I think I was directly able to launch the benchmarks.

Sergio.

[meet...@gmail.com]

Hey Sergio,

I have the 3 environment variables set properly, still I am seeing the issue while running parsec benchmarks. 

Thanks,

Reena

[yinnon...@gmail.com]

hi,

if you have problems with installation - try this (attached)

Yinnon

[Wim Heirman]

Reena,

Looks like there is a bug in our latest benchmarks release where a required file wasn't part of the distribution. Can you copy the attached file to benchmarks/parsec/parsec-2.1/bin (overwriting the original one) and see if this helps?

Regards,

Wim

--
--
--
You received this message because you are subscribed to the Google
Groups "Sniper simulator" group.
To post to this group, send email to snip...@googlegroups.com
To unsubscribe from this group, send email to
snipersim+...@googlegroups.com
For more options, visit this group at
http://groups.google.com/group/snipersim?hl=en
 
---
You received this message because you are subscribed to the Google Groups "Sniper simulator" group.
To unsubscribe from this group and stop receiving emails from it, send an email to snipersim+...@googlegroups.com.
For more options, visit https://groups.google.com/groups/opt_out.

[meet...@gmail.com]

Hi Wim, 

Thanks for your response.

Which file are you referring to? Where can I get it from?

Thanks,

Reena

[Wim Heirman]

It's attached to my last email, unless some email filter removed it?

-Wim

[meet...@gmail.com]

Hi Wim,

I cant see any attachment with your last email :(

[Wim Heirman]

You should be able to download it from here:

http://users.elis.ugent.be/~wheirman/atuin/tmp/parsecmgmt

Regards,

Wim

[meet...@gmail.com]

Thanks for your help Wim! The new parsecmgmt file works, I can run parsec benchmarks now.

I have a question for you though, I tried running the blackscholes benchmark on 4 cores (n 4) and in the roi region. I would expect all the 4 cores will be utilized, however the sim.out shows that very few instructions actually got executed on core 0 (IPC=0), why would this be so? I tried streamcluster benchmark too (supposed to be uniform threads) and I see similar behavior with it.

                                     | Core 0     | Core 1     | Core 2     | Core 3
  Instructions                  |      11667 |   79342505 |   79351688 |   79359285
  Cycles                         |   55635304 |   55634033 |   53752176 |   54740198
  IPC                             |       0.00 |       1.43 |       1.48 |       1.45
  Time (ns)                     |   20915528 |   20915050 |   20207585 |   20579022

Thanks!

[Wim Heirman]

That's just how streamcluster works: the first thread is a management thread that does the initialization (pre-ROI), then spawns the worker threads and waits for them to end.

In addition, our Parsec integration is set up such that with -n X, you always get at most X threads (which guarantees that each thread has its own core). streamcluster itself actually sees -n X-1, other applications that have more management threads see even lower -n values. See parsec/__init__.py lines 80-104 for the exact translation being used.

You can also avoid the thread count translation by adding the thread count to the input set. E.g. run-sniper -p parsec-streamcluster -n 4 -i test-8 runs a 4-core simulation and passes -n 8 to the benchmark, of course this causes multiple threads to time-share each core.

Regards,

Wim