Charting server load

[Alex Beamish]

The two tools that come to mind are top and xload. You probably have top,
but xload is an X11 graphical tool that you may not be able to run.

The load average is how many processes are ready to run at any given time.
Depending on the hardware configuration, the limits to speed may be the hard
drive sub-system, the amount of swap, the network or the processor
throughput.

For example, I have two types of CPU in my grid engine -- one is a 1GHz
machine with 1/2G RAM, and the other is a beefier processor (2.5GHz perhaps)
with 2G RAM. Some jobs that take 3-4 hours on the first class machine take
about 10 minutes on the second class of machine, strictly because the swap
is used less because there's more RAM.

And I can see that by running top -- kswapd is taking 30-50% of the CPU, the
task itself is getting about .1% of the CPU, and the rest is idle time,
waiting for the hard drive subsystem. That's a situation where more RAM
would help; unfortunately these motherboards are maxed out.

If you're experiencing loads of 20, 30 and 40, it's time to move some
processes (and probably users) to other boxes. Web pages should load in 1-3
seconds -- more than that (people have little patience when browsing) and
you'll lose your audience. The load on my busiest server (doing both
database and web server duties) peaked at around 8 yesterday while we
processed almost 300 documents on the grid engine, but it settled down to a
reasonable value of 4, and response time for web requests continued to be
reasonable.

You probably need to contact your web provider and ask them when the new
boxes are being rolled out to lower the load. And if they're not being
rolled out, think about finding a new web provider. (I can highly recommend
pair.com -- FreeBSD servers, a great network and great customer service.)

-- 
Alex Beamish
Toronto, Ontario
aka talexb

On 2/1/07, William O'Higgins Witteman <william.ohiggins-H217xnMUJC0sA/PxXw9srA at public.gmane.org> wrote:
>
> Is there an easy, lightweight way to chart the loads on a server?  I am
> one of many on a shared host that is obviously overloaded, and I want to
> see if I can get the provider to acknowledge that the machine needs
> upgrading or dividing.
>
> What I see is high availability (pings come back fast) but occasional
> high latency (40+ seconds to deliver a web page, 60+ seconds to
> establish an SSH connection) and high loads.  When I see slowdowns and
> log into the box, uptime shows load averages in the 20s, 30s, 40s and
> even 50s.  I don't know what those numbers mean, but they are a lot
> higher then I suspect is good for performance of the web sites on the
> machine.
>
> Any advice would be appreciated.  Thanks.
> --
>
> yours,
>
> William
>
>
> -----BEGIN PGP SIGNATURE-----
> Version: GnuPG v1.4.6 (GNU/Linux)
>
> iD8DBQFFwg1mHQtmiuz+KT8RAhEnAKCzxQtpVE6Kp84jiSf22QBTsU6EIACfVpkJ
> Whj2mp2mVcKQ6755TdrU3jE=
> =0SZ0
> -----END PGP SIGNATURE-----
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://gtalug.org/pipermail/legacy/attachments/20070201/a06ffd70/attachment.html>