Newsgroups: comp.parallel.pvm
From: edwin@dutind3.twi.tudelft.nl (Edwin Vollebregt)
Subject: Re: Performance models for PVM on ethernet
Organization: Delft University of Technology
Date: 22 Jan 1996 14:15:09 GMT
Message-ID: <4e065d$521@elektron.et.tudelft.nl>


Hi all,

Last week I asked the following question, and now give a summary of
the results so far:

QUESTION:

EV> Can anyone point me to performance models for PVM on ethernet, such
EV> that I can predict the time required for communication from the number
EV> of messages and their length ?  

EV> We have parallelised a large flow simulation package using PVM on a
EV> network of workstations which are connected by ethernet. We have done
EV> some measurements, mainly with different grid partitioning strategies.
EV> Unfortunately we cannot detect from the data why one strategy is best
EV> in one situation and another strategy in another situation.

ANSWERS:

1) I have used the following model, and it works quite well.

   time to send a message thru Ethernet (or a switch, or a completely
                                         connected direct network) is
   an affine function of the length of a message
   i.e. init time + time per byte * length

There is some blocking that affects the time slightly (150 byte blocks
on an Ethernet), but this has little affect.  Note that this is for MPI
point-to-point communication, although I suspect that both PVM and MPI
give little overhead.

EV> We have considered this standard model too, except for the blocking
EV> effect, but it doesn't work in our case.

2) "Characterizing the resource demands of TCP/IP" by E. Pozetti, G.
Serazzi, V. Vetland and J. Rolia in the proceedings of the HPCN conference
in Milan, may 1995, B. Hertzberger and G. Serazzi (eds).

EV> This article shows that the blocking effect can be large on some
EV> machines. Especially measurements on two HP workstations show a
EV> soothtaw curve: a slightly larger message can be transferred faster
EV> than a shorter message. This effect is not seen between a SUN and an
EV> HP.

3) ftp.mathcs.emory.edu:/pub/vss/empanal.ps.Z might help you.

EV> This report also proposes the affine function alpha + beta * Nbytes

4) Thierry Delaitre (delaitt@cpc.wmin.ac.uk) wrote:

We are working on a performance model of PVM using discrete event
simulation. Our simulation model is being validated. We have been
looking for end-users collaborators. You are welcome to collaborate.

Keywords: 


-- 
 __________________________________________________________________________
|                            |                                 |           |
| Ir. Edwin A.H. Vollebregt  |  Parallel Algorithms Group      |    0---0  |
|                            |  Section of Applied Mathematics |   /|  /|  |
| phone +31(0)15-2785805     |  Delft University of Technology |  0---0 |  |
| fax   +31(0)15-2787209     |               The Netherlands   |  | 0-|-0  |
| edwin@pa.twi.tudelft.nl    |                                 |  |/  |/   |
|                  http://dutita0.twi.tudelft.nl/PA/pa.html    |  0---0    |
|____________________________|_________________________________|___________|
 note the new telephone and fax numbers ...

