[parsec-users] Questions about provided binaries (dedup & freqmine)

Junjie Wu wujj at umich.edu
Wed Jun 2 13:39:59 EDT 2010


Thank you very much Jim. Dedup works for 4GB ram. I never realized that 
it actually need such great amount of ram.

Regarding freqmine, Christian Bienia has sent me a parallel version. 
Thank you for all your help. :-)

Regards,
Junjie Wu

On 14:59, Jim Dempsey wrote:
> Wu,
>
> The hooks calls are made from within the serial regions of your application.
> And they make no thread library calls. Therefore they need not be compiled
> with OpenMP (or pthreads, etc...).
>
> As for memory requirements, what did you observe as the consumption running
> natively on x86? The fact that a program runs on a 2GB system does not mean
> it requires .le. 2GB. The x86 native run have virtual memory demands that
> exceed 2GB while the simulated machine might not have simulated virtual
> memory. Also check the default stack size inside your virtual machine
> environment. If the stack size is excessively large then you might be able
> to reclaim some simulated RAM by reducing the stack size within the
> simulated OpenMP environment.
>
> Another option you have is to simulate a system with 3GB, 4GB,... RAM. If
> your host machine has 2GB physical you should be able to simulate the larger
> virtual system while having your native virtual memory perform paging. This
> may not be as bad as it sounds. Since, for example, if your simulation
> progresses at 1/100th native rate, the paging that occurs (hidden from the
> simulation system) would occur at 100x what the virtual clock time sees.
>
> Jim Dempsey
>
>
> -----Original Message-----
> From: parsec-users-bounces at lists.cs.princeton.edu
> [mailto:parsec-users-bounces at lists.cs.princeton.edu] On Behalf Of Junjie Wu
> Sent: Tuesday, June 01, 2010 6:50 PM
> To: parsec-users at lists.cs.princeton.edu
> Subject: [parsec-users] Questions about provided binaries (dedup&  freqmine)
>
> Hello everyone,
>
> I successfully compiled and ran all the cases natively under x86, but I'm
> now using SIMICS to simulate a sparc structure with solaris 10. The uname -a
> information:
> SunOS abisko 5.10 Generic_120011-14 sun4u sparc SUNW,Sun-Fire
>
> I don't want to manually compile everything inside the simulator, and I
> don't have a sparc machine either. Therefore, I downloaded the provided
> binary from the website. I use parsecmgmt to run my test (bin/parsecmgmt -a
> run -p<case>  -c gcc-hooks -x pre -i<size>  -n<nthreads>)
>
> I have following problems:
>
> Dedup:
> [HOOKS] PARSEC Hooks Version 1.2
> [DataProcess] Memory allocation failed.
>
> The simulated machine has 2GB memory. Is that enough? I didn't have this
> problem when running dedup under x86 (compiled myself).
>
> Freqmine:
> Based on my experience working with x86, I'm pretty sure the provided
> gcc-hooks binary is not a parallel one. The speedups I got was always
> 1.00xx, no matter how many threads I was using. The default gcc-hooks config
> calls openmp config, but that openmp config never specifies the -fopenmp
> flag. I also got confirmation from someone else running the provided
> sparc-solaris binaries. It's trivial to resolve the issue, but since I can't
> recompile it in simulator, I want to request an update on the official
> binary. I don't think a serial gcc-hooks version is intended. Just for
> anyone interested:
> add:
> CFLAGS="${CFLAGS} -fopenmp"
> CXXFLAGS="${CXXFLAGS} -fopenmp"
> to pkgs/apps/freqmine/parsec/gcc-hooks.bldconf will do the trick. Though
> this might not be a nice way.
>
> I'd really appreciate any thoughts on the two problems.
>
> Regards,
> Junjie Wu
> _______________________________________________
> 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