/ Zope / Apsis / Pound Mailing List / Archive / 2003 / 2003-12 / only small performance gain?

[ << ] [ >> ]

[ Pound-current fix / Robert Segall ... ] [ redirecting with pound / Joachim Schmitz ... ]

only small performance gain?
Sascha Ottolski <sascha.ottolski(at)gallileus.de>
2003-12-11 17:05:08 [ SNIP ]
Hi,

I made some more tests with -current, and didn't find what I expected. 
And now I'm quite confused whether I made the wrong assumptions. I 
wanted to find out how much faster two servers behind a pound proxy 
would perform, compared two one, or compared to both, if they are 
connected directly.

What I expected was, that two server behind a pound would almost double 
the performance. What I found is: It doesn't at all :-(

Some info: I did a very basic testing with the apache "ab" program. I 
tried with somewhat dynamic content as well as with downloading a 
picture. Whatever I tried, the results were almost everytime similar to 
this (the testservers are not the fastet, I know...):

BackEnd1 directly:            2,84 pages/second
BackEnd2 directly:            1,96 pages/second
two BackeEnds behind a pound: 3,12 pages/second

(with a picture it looks something like BackEnd1: 80, BackEnd2: 70, 
behind pound: 90)

The gain isn't too impressive, right? I'm now wondering if this is 
expected? Two more notes: My pound config was very basic:

-----------------------------------------------------------------------
ListenHTTP 127.0.0.1,80
ListenHTTP 192.168.0.98,80
User nobody
Group nobody
ExtendedHTTP 1
WebDAV 1
LogLevel 2

URLGroup ".*"
BackEnd 192.168.0.1,8080,1
BackEnd 192.168.0.98,8080,1
EndGroup
-----------------------------------------------------------------------


And another strange thing I've seen: I checked every server alone 
through a pound with only one BackEnd, to see if this would slow things 
down. But it seemed to do vice versa, the numbers were always a bit 
better than those when connecting the server directly. I fact, I can't 
believe this to be true, but at least it would prove that pound as is 
would not be a problem for the performance.


Thanks for any advice,

Sascha


-- 
Gallileus - the power of knowledge

Gallileus GmbH                   http://www.gallileus.info/


Re: only small performance gain?
Robert Segall <roseg(at)apsis.ch>
2003-12-11 17:40:38 [ SNIP ]
It is not unusual to see little improvement if the bottleneck is somewhere 
else.

I must admit I never saw a server that does 2-3 pages per second: on static 
content Apache usually does at least 35-50 pages per second, even on very 
slow hardware (K6/233, 64MB). I suspect you have something else slowing 
things down (DNS reverse lookups for the logs? Very slow network/bad network 
card/bad wiring? Flaky disk?)

How much speedup you can get depends on a lot of factors, including how fast 
is the machine on which you run Pound itself. If you do SSL a hardware engine 
can do wonders, and a multi-CPU machine is a big gain for systems with good 
support for threads (Linux/Solaris rather than *BSD).

Pound can make a single server somewhat faster by reducing the network 
overhead, but the gain is minimal.

Final note: if you run Pound and one (or more) server(s) on the same box you 
may be able to improve the performance minimally by using localhost 
(127.0.0.1) rather than the external network address.
-- 
Robert Segall
Apsis GmbH
Postfach, Uetikon am See, CH-8707
Tel: +41-1-920 4904


MailBoxer