Re: Re: Re: Multicore Calculations
- To: mathgroup at smc.vnet.net
- Subject: [mg60880] Re: [mg60862] Re: [mg60809] Re: [mg60769] Multicore Calculations
- From: Jeff Bryant <jeffb at wolfram.com>
- Date: Sat, 1 Oct 2005 02:55:57 -0400 (EDT)
- References: <200509280541.BAA08344@smc.vnet.net> <200509290941.FAA01101@smc.vnet.net> <200509301038.GAA29349@smc.vnet.net>
- Sender: owner-wri-mathgroup at wolfram.com
Mathematica on Linux-x86-64 supports multithreading on multicore and
multiprocessor systems. You can set the number of threads with the
OMP_NUM_THREADS variable, eg a value of 1 will disable threading and a
value of 4 will use three threads. By default Mathematica uses all
processors (physical or virtual) listed in /proc/cpuinfo. None of this
changes the fact that BLAS level 1 and 2 operations don't parallelize well.
-Jeff
Manuel Schmidt wrote:
> Hi,
>
> this are the informations about my system:
>
> Dual Opteron
> 2G RAM
> SuSE Prof. 9.3
> Mathematica 5.2
>
> The threads do not show up independently on this system (it is a
> cip-pool). On my local computer, I have only one processor (Athlon64)
> with Gentoo (I know, it is not supported) running and there the kernel
> uses 4 processes. Can one set from outside (as user or installer) how
> many Threads a kernel uses?
>
> Thanks,
>
> Manuel
>
> On Thu, 2005-09-29 at 05:41 -0400, Jeff Bryant wrote:
>
>> The results you get depend on a couple of factors. I'll provide a
>>couple comments and perhaps other, more qualified people can respond
>>directly if necessary for additional details.
>>
>>What platform (CPU and OS) are you running on? That will determine if
>>a) we support multiple processors/cores
>>and
>>b) if top shows threads independently of their parent process.
>>
>>Also, vector-vector (BLAS level 1) and matrix-vector operations (BLAS
>>level 2) don't parallelize terribly well, so there will be little
>>benefit to multiple processors for those types of operations, anyway.
>>
>>-Jeff
>>
>>Manuel Schmidt wrote:
>>
>>
>>>Hi all,
>>>
>>>i'm wondering, how one can convince the MathKernel to do a numerical
>>>linear algebra calculation on two processors at the same time in 5.2.
>>>
>>>Am I right, that this should be computed on two processors (if
>>>available):
>>>
>>>vec1 = Table[Random[], {i, 5000000}];
>>>vec2 = Table[Random[], {i, 5000000}];
>>>
>>>While[True,Dot[vec1, vec2]]
>>>
>>>or a big matrix and a vector
>>>
>>>matrix.vector;
>>>
>>>Anyways, I see in top only one kernel running. There is only one thread.
>>>Are there different binaries for the multicore and the singlecore
>>>version?
>>>
>>>Thanks,
>>>
>>>Manuel
>>>
>>>