Newsgroups: comp.parallel
From: prakash%flamingo@uunet.uu.net (Sundeep Prakash)
Subject: Questions on IBM-SPx message latency
Organization: UCLA Computer Science Dept.
Date: Thu, 9 Feb 1995 17:05:55 GMT
Message-ID: <3hbfm8$a0p@delphi.cs.ucla.edu>


 Hi,
   I've been collecting data on message latencies on the SPx, and have
 been getting a wide variation. All I do is ping-pong a single message
 between two nodes (exclusively allocated to me).
 My questions:
 1) What part of the message latency is due to software overheads
    in the sending and receiving processors?
 2) Is this latency fairly constant, assuming I fix message size and
    communication protocol (I've been using MPL with "User Space")?
 3) What part of the message latency is in the switch? By how much
    can it vary? Can it become the major part of the total latency?
 4) Has anyone developed a probabilistic model for the expected latencies
    in a SP like switch? e.g assuming the switch is currently handling
    k Mb/sec, the latencies may be drawn from some pdf? A prediction
    of average latency at k Mb/sec may really not be useful to me, since
    the time taken by synchronization procedures in programs e.g barriers
    do not average message latencies, but take the maximum.
 5) Has anyone developed an SP switch simulator? It need not be very
    accurate i.e exact number of cycles.. but something which captures
    the variation in latencies to some degree.. 

 Any info is welcome..  

 Thanks,

 Sundeep Prakash
 (prakash@cs.ucla.edu)


