Newsgroups: comp.parallel.mpi
From: luis@quipu.electrum.kth.se (Luis Barriga)
Subject: Re: MPI for shared memory machines?
Organization: The Royal Institute of Technology, Stockholm, Sweden.
Date: 10 Aug 1994 16:19:48 GMT
Mime-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Transfer-Encoding: 8bit
Message-ID: <LUIS.94Aug10181948@quipu.electrum.kth.se>

In article <PATRICK.94Aug5094411@walt.CS.MsState.Edu> patrick@walt.CS.MsState.Edu (Patrick G. Bridges) writes:

       LB> Well, it seems like overkill. MPI on top of P4 which is on top
       LB> of SYSV which itself is a mess. 

   I agree. That's why we don't usually do it. We just use MPI on top of
   P4's message passing, which normally uses sockets. This performs quite
   well on workstation clusters. On other machines, such as MPPs,
   ANL/MSU's MPI implementation uses the native message passing
   system. (NX on the Meiko, Paragon, etc, eui or euih on the SP1, CMMD
   on the CM5, etc) This of course, yields even better performance.

Well all these machines are message passing ....

       LB> I am also intererested in high
       LB> performance and I look forward to is MPI/P4 on on top of
       LB> user-level threads that Solaris is offering.

   You mean using the threads to run the different parts of your program
   and using their shared memory to communicate?  THis would seem
   difficult to do, since all of the processes would have to run in the
   same address space, causing all sorts of headaches. Now, having each
   program be multi-threaded and using sockets to communicate would be
   much more doable...

We have different views of shared memory, I think. For you it is
something like system V provides or what can be achieved by several
UNIX processes using mmap(). In my case, shared memory for threads is
the memory within a single UNIX address space. I can't see headaches
since all sync primitives are provided by Solaris. NO sockets are not
feasible for fine granularity processing.

       LB> It seems that an ad-hoc implementation that follows the
       LB> standard would be faster until better times.  

   I don't really follow your meaning...

I meant using the MPI standard and implement myself those MPI
primitives on top of Solaris Multithreading primitives. This seems too
much work so I will use P4 primitives for shared memory instead of
MPI. I've heard that P4 on Solaris uses Sytem V stuff and not
multithreading ... 
--
-------
Luis Barriga				Tel:    +46-8-752 1379
Royal Institute of Technology		Fax:    +46-8-751 1793
Dept. of Teleinformatics                E-mail: luis@it.kth.se
S-164 40 Stockholm                      WWW:    http://www.it.kth.se/~luis
Sweden

