Adrian,


Ah, OK.  I'm familiar with the balanced system concept in general, but I wasn't familiar with the specific formalisms (particularly Amdahl's set of rules) described in that slide deck.  Thanks!


And, yes, I would say I agree with you here - IO500 is specifically an IO subsystem benchmark, not intended to evaluate any other aspect of the system.  It could be *combined* with other data, but collecting that other data is outside the intended scope.  I'd say this is because how to do that general compute benchmarking is a hugely complex and contentious issue, and also because IO500 is useful specifically because it's bounded in its scope.


- Patrick

From: JACKSON Adrian <a.jackson@epcc.ed.ac.uk>
Sent: Saturday, August 10, 2019 11:53:16 AM
To: Patrick Farrell <pfarrell@whamcloud.com>; Puneet Bakshi <b.puneet@iitg.ac.in>
Cc: io-500@vi4io.org <io-500@vi4io.org>
Subject: Re: [IO-500] Why "Amdahl IO no" is not part of IO500 metric?
 
Hi Patrick,

This is a reasonable intro to the balanced system concept:

https://www-conf.slac.stanford.edu/xldb10/docs/xldb4_thu1040_AlexSzalay.pdf

I think the answer really is that the IO500 is specifically an I/O
metric/benchmark, it's not designed to be a whole system thing.

Not to say there isn't merit in collecting whole system data like this,
i.e. something like a combined top500/hpcg/streams/io500, it's just not
what the io500 was designed for.

cheers

adrianj

On 10/08/2019 17:37, Patrick Farrell via IO-500 wrote:
> Puneet,
>
> I am familiar with Amdahl’s Law, but I am not familiar with Amdahl IO
> and cannot find anything with Google.  Do you have a reference for this,
> explaining how it’s computed and what it’s supposed to mean?  It is hard
> to see how the Amdahl’s Law - that speed up from parallelism is limited
> by the serial portion of a program - has any relationship to the I/O 500
> benchmarks, which are concerned with a series of specific and separate
> benchmarks.  The various benchmark numbers are not part of a single
> parallel program, and so do not have the sort of relationship where
> Amdahl’s Law is relevant.
>
> Regards,
> Patrick
> ------------------------------------------------------------------------
> *From:* IO-500 <io-500-bounces@vi4io.org> on behalf of Puneet Bakshi via
> IO-500 <io-500@vi4io.org>
> *Sent:* Friday, August 9, 2019 11:11:47 PM
> *To:* io-500@vi4io.org <io-500@vi4io.org>
> *Subject:* [IO-500] Why "Amdahl IO no" is not part of IO500 metric?
>
> Dear Sir/Madam,
>
> I am Puneet from India and is a student in Indian Institute of
> Technology Guwahati.
>
> I read about IO500 and (Amdahl) Balanced Systems (such as GrayWulf).
> But, after doing some math, I found some of the top IO500 systems have
> very low Amdahl IO number. Why IO500 does not include (and publish)
> Amdahl IO number as one of its metric? Looks like I am missing something
> but I do not know what. Please help me understand this paradox. This
> question just came out as an academic curiosity and not to endorse or
> refute any system. Please help me understand more.
>
> Regards,
> Puneet
> _______________________________________________
> IO-500 mailing list
> IO-500@vi4io.org
> https://www.vi4io.org/mailman/listinfo/io-500
>
> _______________________________________________
> IO-500 mailing list
> IO-500@vi4io.org
> https://www.vi4io.org/mailman/listinfo/io-500
>

--
Tel: +44 131 6506470 skype: remoteadrianj
The University of Edinburgh is a charitable body, registered in Scotland, with registration number SC005336.