Path: utzoo!utgpu!news-server.csri.toronto.edu!mailrus!cs.utexas.edu!uunet!convex!convex1.convex.com!hamrick
From: hamrick@convex1.convex.com (Ed Hamrick)
Newsgroups: comp.sys.super
Subject: Re: What's a super?
Message-ID: <102750@convex.convex.com>
Date: 3 Jun 90 21:25:28 GMT
References: <10305@batcomputer.tn.cornell.edu> <395@garth.UUCP>
Sender: usenet@convex.com
Organization: Convex Computer Corporation; Seattle, WA
Lines: 31

In article <395@garth.UUCP> fouts@bozeman.ingr.com (Martin Fouts) writes:
>In article <10305@batcomputer.tn.cornell.edu> kahn@batcomputer.tn.cornell.edu (Shahin Kahn) writes:
>
>   A supercomputer is a machine whose purchase is a purely political decision!
>   (smiley here).
>
>There is a lot of truth in your statement, but it isn't restrictive
>enough.  A lot of nonsupercomputermachines (sorry) are also bought for
>purely political reasons....

A particularly common example of this phenomenon is the practice of procuring
supercomputers based on single-job performance instead of throughput (based
on representative workload). 

I'm sure that many of you are familiar with large Cray sites with hundreds
of users with an average job turn-around time (for 1-hour jobs) of a week.
Even though the workload at these sites is dominated by lots of relatively
short jobs, hardware upgrades (and the justification for more money) is
almost always based on single job performance (how many times can you say
Grand Challenges? 8-).

Of course, for any broad generalization there exist specific counterexamples.
However, I challenge those readers of this newsgroup from the National
Supercomputer Centers et. al. (e.g. NASA Ames NAS) to post computer
utilization statistics such as median job turnaround time (by job run
time), number of jobs (by job run time) and other similar data.  I've
noticed that many of you seem to have plenty of spare time (still waiting
for your jobs to run? ;-).

Regards,
Ed Hamrick