[httperf] Connection Resets - need some description

Martin F. Arlitt httperf@napali.hpl.hp.com
Wed, 12 Sep 2001 16:17:56 +0000


Deepak

the server will issue a RST if a client sends a TCP packet
(typically a SYN) to a port that does not have an
application listening on it.  in your case, one possibility
is that the Web server became overwhelmed and stopped
listening for requests (or even crashed).   I would recommend
using netstat to provide yourself with more information.
the -s option provides statistics on numerous conditions,
including:

bash-2.04$ netstat -s
tcp:
        10 connect requests dropped due to full queue
        637 connect requests dropped due to no listener

in this example, netstat reports that 10 connection requests
were dropped because the listening application(s) were not
accepting requests quickly enough and the queue overflowed.
in the second line netstat reported that 637 connection requests
were dropped because there was no application listening for
requests on the corresponding port(s).  In this case, RSTs would
have been issued to the clients.  Note that the numbers reported
by netstat are typically cumulative (from the time the machine
was last rebooted), so you will need to run netstat before and after
each test to determine how many requests were dropped.

In my opinion the problem is with the application, but
I am not familiar with the ones you are using so I can't
tell you which parameters to adjust.

Martin

Deepak Hegde wrote:

> Hi,
>
> In the tests that I run using httperf, I get the
> connection resets. Can some one help me understand the
> meaning of connection resets in the context of
> HTTP/httperf ?. The (load generating) client & web
> server are running under Linux on high-end machines
> connected by GBits/sec network. httperf generates the
> HTTP requests to an EJB application which takes more
> time to run and return results (than a simple HTTP
> request).
>
> Both the web server and the application server (IBM
> WebSphere) are configured to support large number of
> connections.
>
> I guess something in the pipe ay be limiting the
> connection queue lengths and thereby resulting in
> closure of TCP connections. The SYN cookies and SYN
> backlogs are enabled to allow queing of TCP connection
> requests.
>
> Which Apache/WebSphere/OS parameters should be tuned
> to avoid connection resets during httperf tests ?
>
> Any thoughts ?
>
> Thanks,
> Deepak
> ------------------------------------------------------
> httperf --hog --client=0/1 --server=linux01 --port=80
> --rate=100 --send-buffer=4096 --recv-buffer=16384
> --num-conns=1000 --num-calls=1
> Maximum connect burst length: 1
>
> Total: connections 1000 requests 1000 replies 591
> test-duration 25.258 s
>
> Connection rate: 39.6 conn/s (25.3 ms/conn, <=767
> concurrent connections)
> Connection time [ms]: min 735.9 avg 8451.9 max 23538.7
> median 5910.5 stddev 6126.9
> Connection time [ms]: connect 4.7
> Connection length [replies/conn]: 1.000
>
> Request rate: 39.6 req/s (25.3 ms/req)
> Request size [B]: 93.0
>
> Reply rate [replies/s]: min 11.2 avg 22.4 max 36.0
> stddev 11.4 (5 samples)
> Reply time [ms]: response 7720.1 transfer 725.7
> Reply size [B]: header 332.0 content 15071.0 footer
> 2.0 (total 15405.0)
> Reply status: 1xx=0 2xx=591 3xx=0 4xx=0 5xx=0
>
> CPU time [s]: user 0.37 system 24.76 (user 1.5% system
> 98.0% total 99.5%)
> Net I/O: 355.6 KB/s (2.9*10^6 bps)
>
> Errors: total 409 client-timo 0 socket-timo 0
> connrefused 0 connreset 409
> Errors: fd-unavail 0 addrunavail 0 ftab-full 0 other 0
>
> __________________________________________________
> Do You Yahoo!?
> Get email alerts & NEW webcam video instant messaging with Yahoo! Messenger
> http://im.yahoo.com
>
> --
> To unsubscribe: echo unsubscribe httperf | mail majordomo@linux.hpl.hp.com


--
To unsubscribe: echo unsubscribe httperf | mail majordomo@linux.hpl.hp.com