Hi


I plan to implement Pound for our loadbalancing infrastructure, and i
have a question.


How does Pound behave if i shut down de HAport on a backend?

A) it doesn't send any new connection to that backend, but the existing
ones are still served
B) It cuts all connections with that backend

I plan to implement a webserver load aware daemon which lets a port open
until the limit is reached then it closes that port. If the load is back
to normal it opens the port again.


[...]

On Tue, Sep 18, 2007 at 03:49:33PM +0200, Nagy Gergely wrote:[...]
Some sidenotes to your questions.

First, be very very cautious with this. It may easily happen that no backend
servers will be available for a few seconds, since pound checks the backends
periodically. So a server could be awailable and it must wait 0-999msecs for
pound to notice it's back in business. This can happen at the same time for
many servers at the time, and shutting down a server increases the load of the
others, because the client demand won't decrease while the number of available
servers will be lower, and this will result in an increased load at the
available
backend, which will be reported N/A by your monitoring tool...

Secondly, IIRC pound has a balancing algorithm that takes into account the
reply time of the servers, and it manages the servers' weights according to
this parameter on the fly. I'd recommend checking out this method first.


Sincerely,

Gergely Czuczy
mailto: gergely.czuczy(at)harmless.hu
[...]

Yes, i was waiting for an answer like this :)
I didn't consider this issue.

I could also enable an at-least-one communication on my daemon if all
backends are overloaded, but thats a hack, which still could have bad
logics behind it.

How do i use/enable this reply time algorithm, or is it enabled already?

We are in the testing phase, and have set up 2 ucarp enabled servers for
pound, and will be using 4 backends. 

In the actual test setup 1 backend will serve the content-manager
interfaces(admin-pages), and will be also a low priority backend, the
other 3 servers will serve the main backend function, but the dynamic
content comes from the content-manager server.

I'm open for any suggestions we should try/implement in a similar
ha-loadbalacing setup.


Thanks

Greg



2007. 09. 18, kedd keltezéssel 16.07-kor Gergely CZUCZY ezt írta:[...][...]

On Tue, Sep 18, 2007 at 04:43:51PM +0200, Nagy Gergely wrote:[...]
Unfortunately it's in pound(8), so RTFM please. Look for DynScale.
[...]
Another sidenote:

I could get out a bit more optimal performance for fbsd+carp+pound then
from a linux-way implementation. Another definite pro of fbsd that it has
accf_http[1] and kqueue(), which both can be a blessing under heavy load.
And also, pound can be made to use accf_httpd by inserting a single line
or so into the source to the appropriate place. It will have much better
response for a few schemes of attacks.

Also pf was better at defending the box against lower-then-HTTP level
attacks, and can be more easily configured for pound's characteristics.
Like you can define every piece of timeout for every rule, such as having
different timeout-characteristics for public and internal connections, which
can also be a very good idea, since pound doesn't do keepalive to the backends,
so it eats up availabe source ports on high load.

Please note, these are just my experiences while i was testing a pound-based
setup, and I could have made some mistakes.

[1] http://www.freebsd.org/cgi/man.cgi?query=accf_http

Sincerely,

Gergely Czuczy
mailto: gergely.czuczy(at)harmless.hu
[...]

Gergely CZUCZY wrote:[...]
Would just like to check - is this the feature I've seen which must be 
explicitly enabled?
Or does it do this adaptive load balancing (baseed on backend reply 
time) by default?
thanks,
Alex
[...]

Found it,

man pound is open on a terminal for me, but i didn't see that.

Also considering the BSD variant, i can learn into it.

Greg


2007. 09. 18, kedd keltezéssel 16.53-kor Gergely CZUCZY ezt írta:[...][...]

Alex Hunsley wrote:[...][...][...]
Please ignore my question above, just remember that the setting is 
called DynScale!
lex