Network Charteristics and Applications
Scott Shealy
sshealy at asgnet.psc.sc.edu
Fri Jan 4 13:21:39 PST 2002
I know the question I am about to ask has the usual answer "It depends" ....
but I was hoping that some people could answer with some specific experiences.
The Question... Which specific parallel applications/algorithms/problem
classes benefit significantly from bandwidth increases,decreased network
latency or a combination of both?
I am trying to classify applications/algorithms/problem classes by wether
they would benefit significantly from network upgrades with the following
options
100MB switched running TCP/IP
n channel-bonded 100MB switched running TCP/IP
Gigabit ethernet switched running TCP/IP
Myrinet.
I know the first three solutions don't offer much in the way of latency
improvements. So..
Which applications/algorithms/problem classes scale no matter what.?
Which applications/algorithms/problem classes scale with increased bandwidth?
Which applications/algorithms/problem classes scale with decreased latency?
Which applications/algorithms/problem classes scale with decreased latency
and increased bandwidth?
I have read alot of theory on this and read a bunch of stuff by
Foster..etc.... and can make statements like an application that uses
frequent short messages will benefit from a lower latency network. I can
write formulas that use Tc and Ts and pull out ratios out the wazoo... But I
am looking for real world applications that people have experience with and
can share details with or infer some of the ratios for me...
Also if you know of applications/algorithms/problem classes that don't scale
at all that would be of interest too...
Also if anyone knows a general reference that discusses this... that would be
of great interest too.
Thanks for any experience you are willing to share...
Scott Shealy
More information about the Beowulf
mailing list