[9fans] message passing.. sci programming

[9fans] message passing.. sci programming

[Nehal N. Desai]

hi,
has anyone looked into using message passing on
plan9.  We are looking into building a 
smallish plan9 cluster -- between 128 and 512 processors
that we would like run some
physics codes on (eg. weather modeling, QCD,etc)...
right now. those most codes here use MPI or OpenMP.
but is there is a better way that uses the plan
9 architecture in a more optimal (scalable) way.  

nehal

Re: [9fans] message passing.. sci programming

[Nehal N. Desai]

sorry for the elliptic (and incorrect)  nature of the previous email.
here's a corrected one.. 

stupidly yours,
nehal 
> 
> hi,
> has anyone looked into using message passing on
> plan9.  We are looking into building a 
> smallish plan9 cluster -- between 128 and 512 processors
> that we would like to run some
> physics codes on (eg. weather modeling, QCD,etc)...
> right now, most codes here use MPI or OpenMP.
> but is there a better way that uses the plan
> 9 architecture in a more optimal (scalable) way.  
> 
> nehal
> 

Re: [9fans] message passing.. sci programming

[Andrey A Mirtchovski]

I was looking at the 9p protocol for implementing some sort of message
passing in my project.. i had to abandon it for lack of time though...

depending on how close one keeps to the 'all is file' paradigm 9p could
really save the day :)

andrey

On Wed, 28 Mar 2001, Nehal N. Desai wrote:

> sorry for the elliptic (and incorrect)  nature of the previous email.
> here's a corrected one.. 
> 

Re: [9fans] message passing.. sci programming

[Jim Choate]

I'm also curious about such things. Please keep me in the loop...

http://einstein.ssz.com/hangar18

On Wed, 28 Mar 2001, Nehal N. Desai wrote:

> hi,
> has anyone looked into using message passing on
> plan9.  We are looking into building a 
> smallish plan9 cluster -- between 128 and 512 processors
> that we would like run some
> physics codes on (eg. weather modeling, QCD,etc)...
> right now. those most codes here use MPI or OpenMP.
> but is there is a better way that uses the plan
> 9 architecture in a more optimal (scalable) way.  
> 
> nehal
> 

Re: [9fans] message passing.. sci programming

[presotto]

[-- Attachment #1: Type: text/plain, Size: 464 bytes --]

What semanitcs are you looking for with the messages?  If it's
just n to 1 with in order delivery and message boundaries, you
can just use pipes bound into the file system.  The reader does
the bind and then exports (via exportfs) his name space to every
system that wants to send him messages.  Each party can do the same
to affect 2 way communication.  No special tools necessary.

Of course, you have to provide your own pickling/marshaling routines...

[-- Attachment #2: Type: message/rfc822, Size: 1931 bytes --]

From: "Nehal N. Desai" <nehal@acl.lanl.gov>
To: 9fans@cse.psu.edu
Subject: [9fans] message passing.. sci programming
Date: Wed, 28 Mar 2001 21:24:12 -0700 (MST)
Message-ID: <200103290424.VAA06811@fred.acl.lanl.gov>

hi,
has anyone looked into using message passing on
plan9.  We are looking into building a 
smallish plan9 cluster -- between 128 and 512 processors
that we would like run some
physics codes on (eg. weather modeling, QCD,etc)...
right now. those most codes here use MPI or OpenMP.
but is there is a better way that uses the plan
9 architecture in a more optimal (scalable) way.  

nehal

Re: [9fans] message passing.. sci programming

[Ronald G Minnich]

On Thu, 29 Mar 2001 presotto@plan9.bell-labs.com wrote:

> What semanitcs are you looking for with the messages?  If it's
> just n to 1 with in order delivery and message boundaries, you
> can just use pipes bound into the file system.  The reader does
> the bind and then exports (via exportfs) his name space to every
> system that wants to send him messages.  Each party can do the same
> to affect 2 way communication.  No special tools necessary.


I've been wondering about this. Does anyone know how much you lose in
bandwidth over raw IL or TCP links?


ron