SEO

I'm running AIX 5.1 on a set of p630's as mail servers, using postfix,
and amavisd-new as the content-filter to call nai's uvscan as the virus
checker.

We've implemented WLM and multiple postfix instances to ensure that we
get a decent performance on various types of email passing through the
system, and one of which has been assigned a class (Call it bulkl) and a
subclass (Call it uvscan).

bulkl has a max CPU of 35% and a hard limit of 40%.
uvscan has a max CPU of 90% and a hard limit of 90%.
(There are 2 other subclasses, one for each postfix instance [before &
after the content filter] in this path.).


OK. When mail is flowing the 1st postfix accepts the email, the content
filter is called that save sthe email to a ramdisk, then uvscan is
invoked to check the email, after which it's sent to the seocnd postfix
instance for delivery to the recipient.

HOWEVER. I've discovered that the virus checker appears to be a
bottleneck in the system when WLM is enabled. Now I could accept that IF
as designed, the uvscan process was getting 90% of 40% (i.e. 36%) CPU,
but it's only getting anywhere from 8% up to a peak of 25%.

This means that at ~300ms to check an email I SHOULD be getting on a 4
way p630 about (12 emails * 36%) per second through the virus checker.
(About 15,500 per hour). However I never get anything above about
4000/hour...

So far I've checked ramdisk contention & everything else I can think of.
If I get a run of email going through and then run side by side two
copies of uvscan against a single email I get about 350ms when running
as root (In the default class), and 55000ms (i.e. 55 seconds) when
running in the class bulkl.uvscan.

Anyone got any ideas why I don't get my defined CPU utilisation? I
should mention that it ONLY appears to hit short running programs (i.e.
a BIT hit at program start). Programs that are long running get their
allocated CPU fine, and don't exhibit this problem.

What it really looks like is I'm getting a limit of the number of
exec()'s per second... About one per second. No matter what.. If I
change the max CPU from 40% to 25% it made no difference... Same
throughput...

Anyone?

TIA

Hamish.

25% usage on a 4-cpu system means (or can mean) 1 cpu is 100%busy

r u sure your application can use more than 1 cpu?

Holger van Koll wrote:
> 25% usage on a 4-cpu system means (or can mean) 1 cpu is 100%busy
>
> r u sure your application can use more than 1 cpu?

Very sure... It's actually about 75 simultaneous copies of NAI's uvscan
for AIX. The 25% is peak usage according to nmon. Actual average is more
like 12%... Turn off WLM and it all works fine.

Hamie, IMHO WLM and Performance are not too terms that bode together
well ... Does your system become so overloaded that it hangs ? or one
process hogs all the cpu time leaving other processes waiting for cpu
? if the answer to those 2 questions is no, then leave WLM off. It
should only be used as a work around to a badly specc'd server for
workload. If on the other hand you get all 4 cpu's constantly on >
85%, then good luck with your config or buy faster / more cpu ...

Rgds
Mark Taylor

mark taylor wrote:
> Hamie, IMHO WLM and Performance are not too terms that bode together
> well ... Does your system become so overloaded that it hangs ? or one
> process hogs all the cpu time leaving other processes waiting for cpu
> ? if the answer to those 2 questions is no, then leave WLM off. It
> should only be used as a work around to a badly specc'd server for
> workload. If on the other hand you get all 4 cpu's constantly on >
> 85%, then good luck with your config or buy faster / more cpu ...
>

None of the above. The box is well speced (in fact there are 5x 4
processor p630's (@ 1.4GHz IIRC)). The WLM is there because it's a
shared box, and the people who use it will get the performance they have
paid for.

What we don't want is for them to get BETTER performance & so develope
an unreasonable expectation that this is becuase of the wonderful code
etc rather than getting a free ride that shall come to a halt at some
stage. nor do we want them to choke the CPU for other people on the box
who have paid to have the headroom.

WLM appears to work fine (We use it for limiting CPu uage by things like
Patrol, and for boxes that have multiple instances of note on them)
until you get a lot of short lived processes that have sub-second
executaion times at which point it seems to roll up its toes & choke the
performance down to a constant 10-15% CPu usage no matter what the
settings...

H

> Rgds
> Mark Taylor