tag:blogger.com,1999:blog-6569681.post113711435391717660..comments2024-03-24T10:38:16.997-07:00Comments on Geeking with Greg: Power, performance, and GoogleGreg Lindenhttp://www.blogger.com/profile/09216403000599463072noreply@blogger.comBlogger2125tag:blogger.com,1999:blog-6569681.post-1137712063895564192006-01-19T15:07:00.000-08:002006-01-19T15:07:00.000-08:00Thanks, John. I have seen the Google Cluster Arch...Thanks, John. I have seen the Google Cluster Architecture paper. I have a writeup on it if you're interested:<BR/><BR/><A HREF="http://glinden.blogspot.com/2004/04/google-cluster-architecture.html" REL="nofollow">http://glinden.blogspot.com/2004/04/google-cluster-architecture.html</A><BR/><BR/>I'm not sure I see how that paper supports your statement that Google's cluster is CPU bound. I don't see anything in the paper that discusses the cost of disk I/O.<BR/><BR/>The closest the paper seems to get is on page 4 where they compare single proc boxes to more expensive multi-proc boxes and conclude that they prefer the single proc boxes for their cluster since it is 1/3 as expensive and has x3 the RAM.Greg Lindenhttps://www.blogger.com/profile/09216403000599463072noreply@blogger.comtag:blogger.com,1999:blog-6569681.post-1137697479783913722006-01-19T11:04:00.000-08:002006-01-19T11:04:00.000-08:00Greg and b7j0c - read this paper for the explanati...Greg and b7j0c - read this paper for the explanation of why CPU matters.<BR/><BR/>http://www.barroso.org/publications/ieeemicro_google.pdf<BR/><BR/>Logically, if they've managed to parallelize it so well, they've pretty much managed to transform it into a more CPU bound problem...John Khttps://www.blogger.com/profile/14471505239458914968noreply@blogger.com