[Beowulf] SC13 wrapup, please post your own
Lux, Jim (337C)
james.p.lux at jpl.nasa.gov
Wed Nov 27 12:21:19 PST 2013
From: "prentice.bisbal at rutgers.edu<mailto:prentice.bisbal at rutgers.edu>" <prentice.bisbal at rutgers.edu<mailto:prentice.bisbal at rutgers.edu>>
Reply-To: "prentice.bisbal at rutgers.edu<mailto:prentice.bisbal at rutgers.edu>" <prentice.bisbal at rutgers.edu<mailto:prentice.bisbal at rutgers.edu>>
Date: Wednesday, November 27, 2013 11:58 AM
To: "beowulf at beowulf.org<mailto:beowulf at beowulf.org>" <beowulf at beowulf.org<mailto:beowulf at beowulf.org>>
Subject: Re: [Beowulf] SC13 wrapup, please post your own
Since you work in F1...
The guys from Intelligent Light (http://www.ilight.com/) gave a presentation at SC13 on how they helped both Red Bull F1 and Zipp wheels (http://www.zipp.com) with HPC. I think both of these examples show how easily quantifiable ROI can be in certain scenarios. It also shows how you can use different methods to calculate ROI.
In the Red Bull F1 case, they made it easier and quicker for Red Bull to analyze the results of their CFD simulations. Obviously, if you can calculate the # of aerodynamics simulations they were able to analyze per week before and after, you can figure out the value based on the cost of labor for the aerodynamics experts, and then calculate the value of this increase in productivity. The ultimate measure, though, would be to measure the increase in co-sponsor dollars/prize-winnings after using this new workflow.
But this kind of thing is hard to really quantify. It's like trying to quantify the value of a square foot of laboratory space (or office space). You don't really have a nice relationship for productivity: someone will argue that had they had less computational resources, they would have come up with better analytical solutions. After all, the atomic bomb was designed with slide rules, and the computers used for trajectory calcs for Apollo weren't a whole lot better. One will hear exhortations from managers to "do more with less, in these straitened economic times", and most of the time, people will do this (work hours grow, but salaries do not).
(at least until an alternative for the employee arises.. If they can go elsewhere, they will)
It's like those "studies" that show that open plan offices are more cost effective. What it really means is that you have to buy or rent fewer square meters per employee because you aren't procuring walls, hallways, and doors. Even more extreme is "hoteling", where you don't even have cubes. The problem is that "people productivity" for knowledge workers is as hard to measure as HPC ROI. You need many years of data to know if you are working better or worse; but the immediate cost reduction is obvious. So you might be able to save money on a per employee basis, but get less total "value" out of them.
It's only if the alternatives are equally viable that you can do a head to head comparison: A room full of 1950s era "computers" with their Marchand calculators against an old IBM PC doing the same calculations.
How do you quantify (in an absolute, accounting/book-keeping sense) the "value" of running twice as many simulations? There are very few HPC situations where you can draw a curve that has value/additional run. Usually, the number of runs is set either by a resource limit (wall clock time, money, etc.), or by some aspect of the underlying randomness (simulating at a finer scale doesn't provide information because the underlying phenomenon has a larger scale), or the problem becomes numerically unstable or ill conditioned (which is really a "precision" issue). But there are engineering problems where "doing better", particularly with one aspect, can be shown not to improve things. (This can all be formalized with things like Cramer-Rao bounds and the Fisher Information Matrix). You can improve the signal processing in a single frequency GPS receiver to your heart's content, but the accuracy will be limited by the uncertainty of ionospheric propagation.
What is the economic value of getting weather predictions that are better? Over many years, one can figure out statistics that say how much it saves to know where the hurricane landfall will be with higher accuracy: the reduced costs of evacuation, for instance. But even in those situations, the estimates are a bit fuzzy, with lots of externalities.
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Beowulf