[Beowulf] Re: Estimating cluster power consumption - more on I/Issues / Mr. Hahn

Douglas Eadline deadline at clustermonkey.net
Wed Dec 21 15:06:57 PST 2005


Do you have some actual performance data to support these examples?
(stealing a few cycles here and there is quite a bit different than
degrading performance)


> On Dec 20, 2005, at 8:55 PM, Mark Hahn wrote:
>>> extraneous video, I/O, drives and other bus linked hardware and
>>> features, likewise ... improves performance. (I notice that some
>>> performance system builders also fail to "disable" many built-in
>>> features in CMOS BIOS setup ... thus unknowingly degrading
>>> processor / bus performance.)
>> I'm a DIY/minimalist myself, but have never been able to measure
>> any real benefit.  what are these bios features or extraneous drivers
>> that degrade performance?  saving some KB is not a bad thing, but...
> Well, all extraneous I/O of almost any unneeded type should be
> disabled, regardless:
> Example: if the nodes are to be connected via 1000baseT (PCI / PCI-
> extreme / PCMCIA plugin add on cards) and onboard built in LAN
> connectors are 100baseT and thus, not used then it should be obvious,
> but sometimes overlooked, that processor bus communications and
> performance are enhanced by disabling these extra LAN connections.
> Not only do these connections "steal" CPU cycles via interrupt IRQ
> polling, having the CPU spend time generating a default networking
> protocol on the built in LAN chip via this connection certainly does
> "steal" a whole lot of CPU time. (On a higher level, multiple
> protocols on the same LAN connections should likewise be removed or
> defeated.)
> Example: Some nodes may have built in audio features. Although these
> may not prove to be net performance detractors in some cases,
> reliability may be enhanced by disabling these aka the KISS
> principle. (Consider: Newbe SysOps may wish to play MP3 files while a
> node is active.)
> Example: depending on the CMOS / BIOS maker, the built in serial and
> parallel ports actually can steal CPU cycles if not disabled ... even
> on very modern x86 systems, these interrupt IRQs are still "polled"
> for activity unless disabled. (Note that some CMOS / BIOS routines
> that may indicate "disabled" in the setup are actually simply
> "blocked", the CMOS / BIOS may still be examining or polling the
> interrupt / IRQ. For those purists concerned about such CPU cycle
> theft, the BIOS might have to be completely rewritten or a substitute
> made to maximize CPU performance. [ala The Bill Gates Syndrome = what
> does IRQ 11 actually do?] )
> Ed Karns
> FireWireStuff.com
> IndustrialComponent.com
> USBStuff / FireWireStuff / WireLessStuf / FiberStuf ... and much more
> http://industrialcomponent.com/contact.html
> _______________________________________________
> 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