/ Zope / Apsis / Pound Mailing List / Archive / 2005 / 2005-09 / Performance problems

[ << ] [ >> ]

[ Re: [Pound Mailing List] changing conf without ... ] [ Problem with Servlet Redirects / "Sarah ... ]

Performance problems
Andrew Hughes <ahughes(at)lisasoft.com>
2005-09-05 06:06:58 [ FULL ]
Hi All,

I've been looking for a loadbalancer to use for high throughput and
failover capabilities. Pound certainly seems like the answer, however I
am having some performace problems with it:

Overview: Incoming request are load balanced (this offers service level
redundancy), the service places several (5) subsequent http
(loadbalanced) requests to share the actual work across several
machines.

I have been comparing the performance of pound against an apache
rewrite and the differences are significant. Removal of all processing
logic and returning a binary and a text file was used to measure the
"no brains" raw delivery of pound vs apache rewrite, with results
below: 


      
   
    
      
      
      Destination - Apache 
    
    
      
      
      URL
      Pound
      
      ReWrite
      Pound%
      Difference
    
    
      1
      testBinary.gif
      141.2
      254.4
      55.50%
      -113.20 
    
    
      2
      testBinary.gif
      142.2
      290.4
      48.97%
      -148.20 
    
    
      3
      testBinary.gif
      131
      292.2
      44.83%
      -161.20 
    
    
      4
      testBinary.gif
      132.4
      264.7
      50.02%
      -132.30 
    
    
      5
      testBinary.gif
      145.8
      307.4
      47.43%
      -161.60 
    
    
      Avg
      testBinary.gif
      138.52
      281.82
      49.15%
      -143.30 
    
    
      1
      testText.txt
      119.2
      185.4
      64.29%
      -66.20 
    
    
      2
      testText.txt
      111.7
      189.7
      58.88%
      -78.00 
    
    
      3
      testText.txt
      107.5
      185.3
      58.01%
      -77.80 
    
    
      4
      testText.txt
      110.5
      194.2
      56.90%
      -83.70 
    
    
      5
      testText.txt
      112.5
      190.5
      59.06%
      -78.00 
    
    
      Avg
      testText.txt
      112.28
      189.02
      59.40%
      -76.74 
    
    
      Avg
      TOTAL
      125.4
      235.42
      53.27%
      -110.02 
    
  


The two delivery rates of Pound 125.4/sec and Apache
ReWrite 235.42/sec are both very fast delivery rates. However
the difference is significant with pound delivering only 53% of the
traffic apache is doing. 

With html href's and the spauning of "several subsequent http requests"
like we are doing, we require a load balancer with minimal overhead.
Indeed, during a load test of our application we are able to deliver a
sercive of 62/min using apache rewrite. We then
use pound and this drops to a staggering 32/min
(note this is 51%). FYI: apache and pound run on the same machine, but
listen on a different ip address, so Im NOT going through one to get to
the other (that would be stoopid). The machine performing the load
balancing is a Dual Xeon running OpenBSD 3.6.

If anyone has any suggestions for making things faster I would really
appreciate it! I dont require sticky sessions either.


Thanks in advance,


Andrew


ps if there is an irc channel for pound I'd love to know where....
Attachments:  
text.html text/html 13786 Bytes

Re: [Pound Mailing List] Performance problems
Robert Segall <roseg(at)apsis.ch>
2005-09-05 15:17:59 [ FULL ]
I'm a bit confused as to what the numbers represent: is it
requests/second, KBytes/second or what? Also at some point you mention
something per minute - which makes it even more confusing.

Secondly you claim "so Im NOT going through one to get to the other
(that would be stoopid)". How you deliver anything with Pound without
going to some back-end I don't know...

I can assure you that Pound performs as advertised - it is small, lean
and fast. I suggest you explain a bit better how and what you tested and
we'll look at it again.

Final question, slightly OT: you're running OpenBSD on a dual CPU
machine. Much as I like OpenBSD (and I do) that is not a very good idea
- the user-mode threads make it impossible to benefit from the second
CPU. Please try doing your tests either with Linux (kernel threads), or
on a single-CPU machine.[...]

Re: [Pound Mailing List] Performance problems
Andrew Hughes <ahughes(at)lisasoft.com>
2005-09-07 04:38:19 [ FULL ]
Okay, after more and more... and even more testing.....

Robert Segall wrote:
[...]
the server per second, or sometimes per minute.
[...]
not doing something stupid. :)
[...]
balancer crashes, our redundancy kicks in.
Pitty about the CPU, we will take that on the chin.

Now....

Everything has tested quicker via pound! Everything, images, txt files, 
html pages, access to our backend services..... between 140% and 200% 
quicker! Except..... our jsp's run very VERY slow. Even a simple 
"HelloWorld.jsp" runs at half the speed as apache rewrite.

I will raise this in a new thread, but thanks for everyone that read this!

MailBoxer