Good indicator of current load

[Reuben]

Hi, I'm working on a small Perl script to dynamically add or remove
backend nodes for perlbal based on the current load. To do this I need
to extract a number for the current load from perlbal. I've been
looking at the telnet management interface and there's a number of
commands that might give me what I'm looking for, but I'm afraid I
don't understand what they describe well enough to decide which
numbers to use.

Specifically I've been looking at the pending, queues and states
commands and they all seem to produce some useable numbers. Using for
example the "queues" command in a script I can get the average number
of queued connections past x seconds (I think) and I suppose it's a
decent indicator. However deciding based on this how many nodes are
safe to shut down or feasible to boot up seems a bit difficult.

I'm wondering if perhaps anyone with a bit more knowledge of perlbal
could give me a few pointers on how to get a better measure of the
current load.

Thanks in advance for any suggestions.

[dormando]

Hey,

The Perlbal::Manual should go into detail for all of the stats. I'll try
to give some brief notes though:

Most tuned setups should be using backend pooling, connect-ahead, and
queues. "show service blah" will list how many connections are idle for
various reasons.

Once you've hit queues, clients are already waiting and your backends are
past capacity. It's great for smoothing out bumps, but your script should
be watching the average over time and make decisions based on how few idle
resources tend to be available, vs how deep queues get.

Also fwiw the "queues" output are the total clients waiting for a backend
per service at the time you issue the command.