Re: strange performance numbers



On Tue, 18 Feb 2003 17:08:01 CST, Tim Flechtner <timf trdlnk com>  said:

> we have two boxes with identical modeled graphics cards:
> red -
>    a 2.53 Ghz P4, 533 Mhz front side bus with 512 MB DDR RAM running 
> solaris 8 (x86)
> protea -
>    a 2.0 Ghz P4, 400 Mhz front side bus with 512 MB SDRAM running 
> solaris 8 (x86)

> interval   red   protea
> 1000       0.02  0.02
> 100        0.05  0.05
> 10         0.07  1.50
> 
> can anyone suggest why protea's (admittedly a lower power machine) cpu 
> usage would shoot up so dramatically under the heavier load?

I'm willing to bet that somewhere between interval=100 and interval=10,
either the CPU or something else saturates, and you hit a "knee" in the
performance curve - updates are showing up faster than you can dispatch them.

Imagine a line at a bank waiting for a teller - if the teller is able to
average handling clients as fast as they arrive, the line stays short.  But
even if the teller only averages 5 or 10 seconds per client slower than
the arrival rate, very soon you have a *very* long line....

Probable things that bottleneck:  Total CPU, speed of context switches.
Possibly the bandwidth/speed of the underlying flush-to-Xserver code.

Might want to run 'vmstat 5' on each machine and see what is getting pegged
to 100% or other high value for interval=10....
-- 
				Valdis Kletnieks
				Computer Systems Senior Engineer
				Virginia Tech

Attachment: pgpiszJAzuQep.pgp
Description: PGP signature



[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]