[Beowulf] Not all cores are created equal

richard.walsh at comcast.net richard.walsh at comcast.net
Tue Dec 23 09:57:37 EST 2008



John/All, 



It is a useful reminder I guess, but I have to assume that this is 

something that folks on this list are familiar with, No?  State is 

more complicated for parallel work and less deterministic.  As 

these qualities accumulate in the processing of any workload 

a stochastic outcome (timings in this case, but it can also effect 

precision as anyone that has compare vector to scale results 

remembers) is the result. 



This effect is another manifestation of the behavior demonstrated 

by Galton in the late 1800s when he dropped English pence 

through a  pinned board (the pins are the identical cores, the pence 

parrallel processes) ... he got a gaussian distribution.  As the article 

suggests you can reduce the variance by limiting the non-determinism 

in process and that is what eXludus (a Montreal based software scheduling 

company) is doing at the job and process level.  As the number of 

of cores and/or number of processes (virtual or otherwise) grows 

so does the variance in outcomes.  This is a on-die manifestion of 

job skew is it not, and another of the second law of thermodynamics ... 



rbw 


----- Original Message ----- 
From: "John Hearns" <hearnsj at googlemail.com> 
To: beowulf at beowulf.org 
Sent: Tuesday, December 23, 2008 6:54:39 AM GMT -05:00 US/Canada Eastern 
Subject: [Beowulf] Not all cores are created equal 


I'm surprised this has not been flagged up yet. Shamelessly passed on from Slashdot: 
http://www.gcn.com/online/vol1_no1/47765-1.html 
_______________________________________________ Beowulf mailing list, Beowulf at beowulf.org To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf
-- 
This message has been scanned for viruses and
dangerous content by MailScanner, and is
believed to be clean.

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://www.clustermonkey.net/pipermail/beowulf/attachments/20081223/a3f0d3e5/attachment-0001.html>
-------------- next part --------------
_______________________________________________
Beowulf mailing list, Beowulf at beowulf.org
To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf


More information about the Beowulf mailing list