Persistent Connections

[Dmitry Kurochkin]

Hi Jacky.

unjc email <unjc.email at gmail.com> writes:

> Hello there,
>
> I have a question about enabling persistent connections in workload.  The
> following is how I setup the server and robot in the pg file; as shown,
> pconn_use_lmt is set to 1000.
>
>
> Server S1 = {
>     kind = "S101";
>     contents      = [ JpgContent: 73.73%, HtmlContent: 11.45%, SwfContent:
> 13.05%, FlvContent: 0.06%, Mp3Content: 0.01%, cntOther ];
>     direct_access = contents;
>     addresses = [ '25.57.0.10:9090', '25.57.0.11:9090' ]; // where to
> create these server agents
>     http_versions = [ "1.0" ];  // newer agents use HTTP/1.1 by default
>     pconn_use_lmt = const(1000); // Persistent connections - should tune
> this value
> };
>
>
> // Note that this Robot has an undefined request-rate in order to enable a
> // best-effort workload.
> Robot R = {
>
>     kind = "R101";
>     pop_model = { pop_distr = popUnif(); };
>     recurrence = 50%;
>     req_rate = undef();
>     origins = S1.addresses;      // where the origin servers are
>
>     addresses = robotAddrs(authAddrScheme, theBench);
>     pconn_use_lmt = const(1000); // Persistent connections - should tune
> this value
>     open_conn_lmt = 1; // maximum concurrent connections
>     http_versions = [ "1.0" ];  // newer agents use HTTP/1.1 by default
> };
>
>
> I examine the tcp streams of tcpdump output from the client machine, in a
> single-robot test; although "connection: keep-alive" are found in both
> request and respond headers, I see client issue [FIN, ACK]'s every few
> (<10) requests, that is way before 1000 requests they make.
>
>
> GET /w1b7335ec.2b642c95:00000008/t03/_0000413f.jpg HTTP/1.0
> Accept: */*
> Host: 25.57.0.10:9090
> X-Xact: 1b7335ec.2b642c95:00000002 1b7335ec.2b642c95:00020522 0
> X-Loc-World: 1b7335ec.2b642c95:00000008 -1/16703 8351
> X-Rem-World: 1b7335ec.2b642c95:00000008 -1/16703 8351
> X-Target: 25.57.0.10:9090
> X-Abort: -324104509 -1205953971
> X-Phase-Sync-Pos: 0
> Connection: keep-alive
>
> HTTP/1.0 200 OK
> Cache-Control: private,no-cache
> Pragma: no-cache
> Date: Wed, 03 Jul 2013 19:50:22 GMT
> Connection: keep-alive
> Content-Length: 9479
> Content-Type: image/jpeg
> X-Target: 25.57.0.10:9090
> X-Xact: 1b7335e7.5e615116:00000002 1b7335ec.2b642c95:7ffdfadd 0
> X-Rem-World: 1b7335ef.4e10652d:00000008 -1/15998 7999
> X-Abort:  2013317368 2072661844
> X-Phase-Sync-Pos: 0
>

You set the maximum number of open connections per Robot to 1.  This
means that every time a Robot needs to make a request to a server
different from the previous request, it has to close the existing
idle persistent connection.  The persistent connection can only be
reused if the next request happens to target the same server as the
previous one.  Otherwise the existing connection has to be closed to
honor the open_conn_lmt setting.

This would work in case there is just one origin server or a proxy is
used.  For two servers and no proxy you should set open_conn_lmt to 2 at
least.  Or just leave it unset, a Robot should not use more than 2
connections anyway.

>
> I also found the robot machine runs out of ephemeral ports shortly after
> the starrt of the single-robot test.  The ulimit value of the machine
> is 65536.   I am surprised to see this if the persistent connections are
> being used.  FYI, this is non-proxy test.
>

As described above, persistent connections are effectively disabled
(i.e. rarely reused) in your workload.  Clients have to close
connections and open new ones at a high rate.  Hence ephemeral ports run
out.

> 003.04| EphPortMgr.cc:23: error: 4096/8191 (s98) Address already in use
> 003.04| OS probably ran out of ephemeral ports at 25.57.100.2:0
> 003.04| Client.cc:347: error: 4096/8192 (c63) failed to establish a
> connection
> 003.04| 25.57.100.2 failed to connect to 25.57.0.11:9090
>
>
>
> Would you please kindly advise what I might configure wrong?
>

I hope the above helps.

Regards,
  Dmitry

>
>
> Thanks,
> Jacky
> _______________________________________________
> Users mailing list
> Users at web-polygraph.org
> http://www.web-polygraph.org/mailman/listinfo/users