Re: determining host performance for task assignment

On Sun, 14 Mar 2004 [email protected] wrote:

> In other words, you mention the "desirability" of a cpu being a
> combination of many factors --- why not measure that "attractiveness"
> empirically/dynamically by running a small test occasionally ? Better
> yet provide a clean interface for re-running this test / caching its
> result etc. That is what i'm doing but am not too happy with.

there's no one test that will keep anyone happy. That's the rub. That's 
the big problem with the 'information power grid' metaphor -- computers 
are not generators, and you can't measure them with a few constants the 
way you can with a generator. YMMV for all kinds of weird reasons. 

This is one reason there is so much snake oil in HPC. None of the things 
that matter are quantified. Considered that the biggest most public 
quantified list we have, the Top 500, is based on a benchmark that is of 
no value to many of the machines on it. Some argue that it is a bogus 
benchmark. And yet this 'bogus' benchmark drives the purchase of a lot of 
hardware ... 

The only thing that will tell you how you will run on a machine is ... 
your app on the machine. So it won't be a global specmark that tells you 
how well it will work, it's going to be your private number for a known 
machine of known type, based on running on it.

ron