From: Jim Phillips (jim_at_ks.uiuc.edu)
Date: Wed May 18 2011 - 12:48:52 CDT
Just to dispel any misconceptions, rsh or ssh are only used the launch the
processes. This is true for ethernet or Infiniband. The actual parallel
communication over ethernet uses either UDP or TCP sockets depending on
which build of Charm++ is used (which is why some of the released binaries
are labeled Linux-x86_64-TCP).
If you want to increase your performance you are better off getting a
single machine with two GPUs and a faster CPU with more cores than
building a cluster of two cheap machines.
On Tue, 17 May 2011, Darko Stefanovski wrote:
> Hi Francesco,
> GTX460 has 336 CUDA cores and if you go with three of them you will be
> getting 1008 CUDA cores.That is very close to the performance of a GTX590
> which has 1024 CUDAs and approximately the same memory. Now in regards to
> working with two computers with CUDAs, We do have a setup like this here but
> unfortunately working through Ethernet is extremly slow. Part of the slow
> down is associated with the protocol you will be using. I tried both rsh ans
> ssh and two machines working on the same problem as one machine was much much
> much slower. Perhaps we are doing something wrong.
> Best wishes,
> On 5/16/2011 1:45 AM, Francesco Pietra wrote:
>> I have followed previous posts on small-scale GPU-CPU, whre Axel
>> (Professor Kohlmeyer) gave key indications.
>> For a classical-MM-dedicated machine (i.e. single precision
>> calculations), i am trying to spare money (and electricity bill) as
>> much as possible, considering that as a particularly temporary
>> investment (future of integrated GPU-CPU cards?). Thus, I am
>> considering a consumer motherboard based on AMD Phenom II X6 1075T CPU
>> and 2/3x GTX460s.
>> My question is, is there any experience on running NAMD on such
>> graphic cards? If yes, is there any experience on a cluster made of
>> two such machines? (considering that I have no more an electronic
>> engineer helping me).
>> Thanks a lot for advice
>> francesco pietra
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:57:08 CST