Hi,

On Mon, 20 Oct 2003 15:24:40 +0200, Robert Segall 
<roseg(at)apsis.ch> wrote:
[...]

1.)
Possiblity to Loadbalance Layer 3:

IPGroup "10.3.54"
BackEnd1
BackEnd1

How about better thead support for MPs and threading?  Process groups and
CPU affinity and some mechanism to manage the new posix threading model.


On Thu, 13 Nov 2003, Alexander Lazic wrote:
[...]

On Thursday 13 November 2003 23:41, Alexander Lazic wrote:[...]

What exactly do you mean by Layer 3? Should we ignore HTTP altogether?

A more detailed description would be good here.[...]

On Friday 14 November 2003 05:24, Dennis Allison wrote:[...]

In using the system thread library we wanted to avoid exactly that: we were 
hoping that the native implementation would be good enough to take advantage 
of multiple CPUs, etc. BTW - with stock 2.4 kernels Pound benefits 
significantly from MP (tested).

Process groups and CPU affinity are perhaps better handled at the shell level. 
Should they really be part of an application design?

Finally - the dreaded NG thread package: at this point I feel we are all still 
a bit in the dark about what exactly it will be. We have all seen the mess 
RH9 is in this respect, so I suspect we would be better off in waiting for 
the 2.6 kernels and final version of the new libraries before we start 
mucking with it.

Let me know how you see it.[...]

Yup, RH8 and RH9 are a mess... 
I'd like to see something that works under RH9 as I have an immediate
requirement--but that will go away soon.  I am abandoning Redhat in 
favor of Gentoo or the like--I suspect that I'll be trading one set of 
problems for another--but I hope to feel better about system stability,
performance, and stability.

On Fri, 14 Nov 2003, Robert Segall wrote:
[...]

One more thing--about CPU affinity--I think we need to incorporate it to
support Zope.  There's the nightmare scenario where the python global lock 
blocks progress on all processors becasue it cannot guarantee locality.

On Fri, 14 Nov 2003, Robert Segall wrote:
[...]

On Friday 14 November 2003 16:20, Dennis Allison wrote:[...]

We use Gentoo here and so far are very happy with it. Make sure you have 
enough spare disk space and a fast CPU for recompiling the packages...[...]

On Friday 14 November 2003 16:25, Dennis Allison wrote:[...]

That is a python-specific problem - result of tacking-on threads after the 
program was ready. Not an issue in Pound, but rather a nice-to-have in 
Linux...[...]

On Fre 14.11.2003 16:07, Robert Segall wrote:[...]

Yes, sorry but i have used Operas M2 but not configured :-(

1.)
Possiblity to Loadbalance Layer 3:

IPGroup "10.3.54"
BackEnd1
BackEnd1
.
.
EndIPGroup
[...]

Yes because some 'content' don't need to look into the 'HTTP-Content'.

Here are now some more _whishes_ ;-))

2.)
Handle some 'Subpattern' in the session cookie:

Cookie: SESS=$HEXNUM!THE_WANTED_VALUE!$LAST

So that the Session go only to the THE_WANTED_VALUE.

3.)
On Demand reduce the Backends, with such as 'SIGHUP'

So if you change the config from backand 4 such as you say in the
Configfile:

Backend4 IP:PORT:HA graceful

and make a

kill -HUP $POUND_PID

then pound look at the 'cookie' and if there is an Active Cookie-Session
the he give the
request to Backend4 but on 'new' Connects Backend4 dont are in the Lb
calculation included.

I hope i have it right described, if not is there some one which speak
German ;-)
[...]

Nice Task ;-))

How about to look on http://www.pcre.org/
for some Performance wins if
there are some?

al ;-)

On Friday 14 November 2003 20:12, Alexander Lazic wrote:[...]

I'm not clear what this means. Could you please explain the semantics?
[...]

Pound is a HTTP proxy, so I'm not sure I agree. Could it be that you are after 
a generic balancer?
[...]

Good idea - noted towards 2.0
[...]

The idea was raised several times in the past, and every time we found 
problems with it: for example, what if the new configuration defines other 
back-ends and rules altogether?

Unfortunately I don't think there is a good solution for it - for the moment 
we'll stick to the dhcpd-style kill/restart solution. I'm open to other 
suggestions any time.

You may want to look at a possible work-around: use the config with the 
optional HA port, but have a program other than your web server listen to 
that port. Start/stop the program in order to control access to your 
back-end.
[...]

Quite a few people on this list speak German, but I think we should keep the 
discussion in English - it is the lingua franca of the Internet. I am willing 
to consider Latin (the predecessor) as an alternative :-))
[...]

???
[...]

I'm not sure how would that help...[...]

On Mon 17.11.2003 15:10, Robert Segall wrote:[...][...]

You are right but i should be a nice Feautre ;-)
[...][...]

Thanx.
[...]

I think the Admin should know what he/she do. If he kick off the
backend from the 'Session' he/she can do it if they want.
[...]

Hm how about to 'dump' the session-state out and after the start look
into the dump and if you have a session with a backend well load it.

How about this?

Where is the right point to add this in pound, i have see in 'svc.c':

static char *
get_key(GROUP *g, struct in_addr from_host, char *url, char **headers)

i think there can i look into the group for the 'graceful'.

Or is there another point to add it?
[...][...]

I think that is not so easy and therfore interesting ;-)
[...][...]

Hm i have see that many 'Highperformance' Server prefer this lib for
regex or the rx lib i self use the pcre because i like the regex-syn of
perl ;-)

al ;-)