Zufeng Huang wrote:
My steps:
General remarks: Make sure, that you understand the bottlenecks of your
test runs x.1 and x.2. E.g. is one of them slower, because it saturated
available network bandwith, or is it slower, because there are not
enough threads in the AJP connector's thread pool. The best bottleneck
to get is completely saturated CPU.
Your ab runs on a different system?
1, Using a jsp as the target ONE.
1.1, ab ¨Cn 25000 ¨Cc 1000 URL_via_apache
Concurrency Level: 1000
Time taken for tests: 410.105255 seconds
Complete requests: 25000
This is very bad throughput! Using the default thread count on the AJP
connector?
Try with concurrency 1 and then do a couple of runs with increasing
concurrency to study the behaviour.
I would expect latency to go up, especially for trivial dynamic content
and static content, but I would not expect to get a huge decrease in
throughput, as long as you add enough CPU resources for the additional
httpd overhead. You can approximate the relative CPU overhead by
measureing CPU needed with concurrency one (and enough requests to
produce mesurable CPU times).
In general, since latency goes up, you would also need higher
concurrency for the optimal throughput with httpd/Tomcat instead of
Tomcat standalone. In your case I expect concurrency 1000 to be beyond
the optimal point for 1.1 and 1.2. Usually you start by defining some
throughput goal (e.g. in requests per second) and you increase
concurrency until you get the best throughput. If it is above your
requirement, and the response times are good enough, you are done.
1.2,ab ¨Cn 25000 ¨Cc 1000 URL_direct_to_tomcat
Concurrency Level: 1000
Time taken for tests: 12.297322 seconds
Complete requests: 25000
2, Using a html page as the target TWO.
2.1, ab ¨Cn 25000 ¨Cc 1000 URL_via_apache
Concurrency Level: 1000
Time taken for tests: 25.655253 seconds
Complete requests: 25000
2.2, ab ¨Cn 25000 ¨Cc 1000 URL_direct_to_tomcat
Concurrency Level: 1000
Time taken for tests: 24.132493 seconds
Complete requests: 25000
The numbers are so close, that I would expect, that both solutions did
not saturate your CPU and you reached some other bottleneck, like
network bandwith, the ability of ab to handle client requests, or the
number of TIME_WAIT connetions in your tcp connection hash.
Did you use http keep-alive?
1, According to my configurations, apache(2.2.4) has NO advantage against
tomcat(5.5.15) in processing static content.
I'm not sure, if your measurement shows that. But I wouldn't argue for
httpd because of performance reasons. For me it is more the
configuration flexibility, which makes it very powerful in complex
integration scenarios.
2, In processing dynamic content, apache make performance decrease
dramatically.
Performance in terms of latency: yes for fast requests.
Performance in terms of throughput: in my experience no, if you add some
CPU ressources and do good sizing.
It¡¯s the matter of MPM-prefork/worker???
In general prefork should be fine for performance, because it is less
complex than worker. But prefork needs slightly more ressources. So make
sure, that you have enough memory for your more than 1000 httpd children.
FYI:
<IfModule prefork.c>
StartServers 50
MinSpareServers 50
MaxSpareServers 100
Your test run is a little short in order to make startup of 900
additional processes neglectable, at least once your httpd/Tomcat
combination is close to the Tomcat performance. So increase
MaxSpareServers to 1000 and make sure, that you do some test run before
doing the one you are taking your results from.
ServerLimit 2000
MaxClients 1500
MaxRequestsPerChild 0
</IfModule>
------------------------------
maxThreads="1500" minSpareThreads="20" maxSpareThreads="50"
enableLookups="false" redirectPort="8443" acceptCount="300"
Also in your AJP connector?
Are httpd and Tomcat on the same system?
Regards,
Rainer
---------------------------------------------------------------------
To start a new topic, e-mail: users@tomcat.apache.org
To unsubscribe, e-mail: [EMAIL PROTECTED]
For additional commands, e-mail: [EMAIL PROTECTED]
