Web Search Using Small Cores: A paper on Microsoft Bing Architecture

[yarapavan]

http://perspectives.mvdirona.com/2009/09/27/WebSearchUsingSmallCores.aspx

[jbert]

Interesting stuff. Does anyone know easy way to see the instructions-per-cycle on a running (linux) server?

Or is it necessary to instrument each application with hires timers and count instructions between the timer points?

[Freaky → jbert]

I don't know about Linux, but on FreeBSD you can use pmcstat to instrument a running process; see pmc(3) and pmcstat(8).

Linux probably has something similar.

    kldload: can't load hwpmc: Exec format error

Meh, need a custom kernel from the look of it, HWPMC_HOOKS isn't in GENERIC :(

[jbert → Freaky]

Thanks for the pointers.

Looks like the linux/ubuntu equivalent may be 'pfmon', also not enabled in the stock kernel.

I think this is the relevant project page: http://perfmon2.sourceforge.net/man/pfm_get_cycle_event.html

[wmf → jbert]

You might be able to do it with oprofile or the (poorly-named) "perf" tool in the very latest kernels. Or VTune if you want a GUI.

[gcb]

it's a paper on using Atom cores instead of more beefy ones.

The researchers just choose to use Bing for their test data instead of mysql or something else... and their point was that bing was more cpu intensive, so the results would be more solid.