Alteon detail
Anders Berg
andersb at vgnett.no
Sat Aug 19 23:26:08 CEST 2006
On Aug 19, 2006, at 20:14, Poul-Henning Kamp wrote:
>
> Can one of you tell me which state the Alteon was left in,
> I'm going to try to see if I can find out why it hates us,
> and it would be useful to know what metric and weight is
> in effect
I was left in leastconn metric and weight 1. Other Squids weight 5
and 1 Squid in 10.
Thanks for todays testing guys. Great that we got to test some.
I feel we should focus on one thing right now, and that is getting
Varnish to function when we give it all of the traffic. The multiple
Host: testing can wait. We have to find out if it is FreeBSD or
Varnish that can't handle it to good. Getting to know Alteon is
needed. Though more in the context of opening connections, keeping
them alive and closing them. Not so much in the metric that put
traffic there in the first place. One way we can try to even out the
traffic more in a linear way when switching weights is to maybe use a
source-adress hash (sticky sessions) and then use weight. Weight
should then use the number of IP adresses to weight with, not the
mystic leastconn (based on connections). That way we can maybe
control the amount of traffic we push to Varnish more. But I fear we
will not get to much out of balancing on the work/not-working
threshold. Either we are hitting a FreeBSD barrier, or a Varnish
barrier, and we should be able to pinpoint that out with just doing
what we normally do before we give all the traffic to Varnish (weight
30-40 Varnish, 5 on others, then 100% to varnish). The barrier should
show itself somewhere.
If you guys don't diagree we should just focus on switching between
working and not-working traffic patterns next time, and see what we
find. We can probably learn some from looking how Squid does it's
connection handling.
I am gonna teach DES some Alteon commands tomorrow, so he can control
some traffic while I am gone.
I can spend a couple of hours tomorrow throwing some traffic on
Varnish if you guys are up for it?
Anders Berg
> --
> Poul-Henning Kamp | UNIX since Zilog Zeus 3.20
> phk at FreeBSD.ORG | TCP/IP since RFC 956
> FreeBSD committer | BSD since 4.3-tahoe
> Never attribute to malice what can adequately be explained by
> incompetence.
>
More information about the varnish-dev
mailing list