[users@httpd] Apache 2.0 Performance Evaluation - strange results

[Davy De Winter <davy.dewinter@xxxxxxxxxxxxxx>]

Hi all,

recently we've did a performance-evaluation on Apache 2.0.

Our testing machine software:
   - debian sid / unstable with apache 2.0.54 with mpm worker
   - kernel 2.6.11

Our testing machines hardware:
   - dual-cpu machine (2 x AMD Opteron 1.6 ghz, 2048 mb ram and pci-x 
busspeed to note the most important bottlenecks) connected with an 
Avalanche machine emulating client-HTTP-requests via 3 ethernet-cables@ 
1Gbit         (in total, 3 Gbit/s traffic flow is possible in that way).

Our main-configuration
   - MaxClients: 8192
   - ThreadPerChild: varied

We've had to determine how many users per second can download a 2 
MB-file (so only request per transaction) and we wanted to determine the 
optimal number of threads per process supporting the highest number of 
(successfull) transactions/sec. We started with 10 requests/sec and 
after 25 seconds, we augmented this to 20 requests/sec and so on up to 
200 requests/sec (bandwidth becomes a bottleneck). The access-network 
speed of the users is in theory unlimited (in practice, this boils down 
to 110 Mbit/s). So we started with 2 threads per child and doubled the 
number of threads with each new test up to 128 threads (which is the 
maximum supported otherwise we get an error message - no memory left to 
create a new thread). Now some "strange" results:

- 4 threads per process gives the best performance-results: up to 200 
successfull transactions per second of the 2 MB file are realized. 
(eventually, the CPU becomes the bottleneck)
- 8 threads per process gives worst performance: only 130 successfull 
transactions per second are possible (and memory is clearly the 
bottleneck here).
- 2 threads per process or 16 / 32 / 64 / 128 threads per process gives 
almost similar results: +/- 140 successfull transactions/sec are 
supported (memory is the bottleneck - page faults / swapping starts -> 
strange with only one file!)

Can someone explain this why I get that different results, especially 
why exactly 4 threads is optimal. If someone is interested in the 
graphs, I can always mail them as a jpeg-figure.

Thanks in advance!

With kind regards,
Davy De Winter.


---------------------------------------------------------------------
The official User-To-User support forum of the Apache HTTP Server Project.
See <URL:http://httpd.apache.org/userslist.html> for more info.
To unsubscribe, e-mail: users-unsubscribe@xxxxxxxxxxxxxxxx
  "   from the digest: users-digest-unsubscribe@xxxxxxxxxxxxxxxx
For additional commands, e-mail: users-help@xxxxxxxxxxxxxxxx