From: Greg A. Woods (no email)
Date: Fri Aug 01 2003 - 11:13:17 EDT
[ On Friday, August 1, 2003 at 15:04:34 (+0200), J.A.J. van Belkum wrote: ]
> Subject: Re: when things go wrong
>
> I understand that LA is the following:
>
> The average number of processes that the scheduler has to choose from
> when it decides which process is going to run next.
>
> So (simply put) if you have 1 processor with a LA of 2 (so you have an
> average of two processes waiting to be run), and you add a processor....
> One process will go to the new processor and your LA will drop by 1.
Generally speaking that's correct. Note that "load average" must always
refer to a time period over which the "average" was calculated and
traditionally there have been three load averages reported: a 1-minute,
a 5-minute, and a 15-minute.
However "load average" is a very simplistic and often rather useless
measurement of system throughput and capacity, especially when it's
discussed without mention of its exact implementation, including not
only the type and version of the OS, but also the complete hardware
architecture.
- It's not necessarily a bad thing to have a lot of processes waiting
for the CPU(s) -- i.e. a high load average isn't necessarily "bad".
If you can type "uptime" and get a response in a resonable amount of
time then overall the system is still functioning just fine even if it
does have a 1-min or even 5-min load average in the hundreds or
thousands. Things will go well even for the jobs accounting for the
high load average if they're fair and low consumers of CPU. If your
system has a 15-minute load average in the hundreds, and if it stays
there for an hour or more, then you might want to investigate. If you
find nothing wrong (i.e. everything is functioning correctly, just
slowly, and there's simply a lot to do), and if this is a regular
occurrence, then you should consider getting a faster CPU or finding
some way to spread the load over more CPUs.
- Adding CPUs won't always decrease the load average, though it's also
theoretically possible to have a situation where adding just one CPU
will decrease LA fully by half (even in an SMP architecture where no
other changes are made)!
- some operating systems are notoriously bad at keeping a good measure
of their overall load average, even one based on the very simple
description above.
-- Greg A. Woods +1 416 218-0098 VE3TCP RoboHack <> Planix, Inc. <> Secrets of the Weird <>
|
|
|