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

[Josip Loncaric]

Douglas Eadline wrote:
> Ed,
> 
> 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)
> 
>>
>>
>>Well, all extraneous I/O of almost any unneeded type should be
>>disabled, regardless:
>>
>>Example: [...]

A few cycles should not matter on a single machine, but unsynchronized 
system daemons (e.g. polling unused hardware, which may take 10-100ms, 
say, once per minute) can have a huge impact on parallel system 
performance.  Typically, parallel codes have frequent synchronization 
points, so if a system daemon kicks in on *any* node, the *entire* 
parallel code will be delayed by that much.  With many nodes, it becomes 
very likely that *some* node will run a system daemon.  If the daemon 
service time is about as long as the application synchronization 
interval, you could cut performance in half with high probability.

Therefore, I like the KISS principle and normally turn off whatever 
isn't essential.

See the paper "The Case of the Missing Supercomputer Performance" by 
Petrini et al. at http://www.sc-conference.org/sc2003/paperpdfs/pap301.pdf

Sincerely,
Josip