Similar rational arguments apply to why you want to invest into a good compiler. I find though easier to justify these things in terms of the metric money rather than performance metrics. Just HPC people are not that used to use the metric money. In other words it becomes a business decision.
Example : Compiler A, 10% better than Compiler B in terms of performance for 80% of applications on HPC infrastructure of cost 1M USD. Compiler A costs 30K USD. Compiler B is open source (nothing wrong against GCC or Open64). It means you can aproximately reduce your computing infrastructure and electric bill and maintenance cost by a factor of 0.10*0.80=0.08 (8%). Total cost of system is 1MUSD (HW). So you can have an impact on the savings in hardware about 80K USD. That by itself will compensate to pay 30K USD on the compiler. Add to that savings the power savings and maintenance savings in the same percentage for 3/4 years of life of the infrastructure. So 30K saved you 80K on HW+80K on electric bill for 3/4 years and 80K for maintenance of 8% fewer nodes. Total investment 30K , total savings 240K-30K=210K. Finally, with those 210K you hire a good HPC guy to optimize the code for a solid year ;) Joshua ------ Original Message ------ Received: 03:30 AM CEST, 04/06/2013 From: Robin Whittle <[email protected]> To: Beowulf Mailing List <[email protected]> Subject: [Beowulf] Often favorable to hire HPC specialists over more hardware > Parts of the "Roadrunner shutdown" thread developed into a discussion > about the benefits of hiring specialized HPC programmers as an > alternative to spending more money on hardware. > > For the benefit of folks searching the archives, here is one of the > messages in that sub-thread, from Brian Dobbins, which contains many > points of interest: > > http://www.beowulf.org/pipermail/beowulf/2013-April/031271.html > > including: > > Why it may be worth hiring a good developer if there are a moderate > to high number of nodes in the system already - with rough dollar > cost estimates. > > A number of programming- and compilation- level things which can > easily be fixed by a programmer with the right expertise, which > will greatly improve performance, but which might not be known by > programmers who are primarily interested in science. Also, how > common they are. (I wonder: if this is true, then perhaps half or > more of HPC cluster time is being wasted on this stuff, in > environments where many types of jobs are being run. I figure > the people who devote clusters to one or a few programs probably > invest in writing and tweaking those programs well.) > > Having a skilled programmer to ensure to that the results are > correct, not just that there are results at all, or results > faster than before. > > "So that's my lengthy two cents in defense of why it's /very often/ > favorable to hire HPC specialists over more hardware . . ." > > I know many people tend to like short chatty messages on mailing lists, > but I like ones like this, for which Brian's apology: "(PPS. Sorry for > the length!)" absolutely does not apply. > > - Robin > > _______________________________________________ > Beowulf mailing list, [email protected] sponsored by Penguin Computing > To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf _______________________________________________ Beowulf mailing list, [email protected] sponsored by Penguin Computing To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
