Re: Performance question

On Tue, 15 Jun 1999, Dirk Moerenhout wrote:

]First of all your settings seem to be way over the top. There is no reason
]to set tcp_conn_req_max_q and tcp_conn_req_max_q0 to such high values.

Unless you are fearing SYN floods, where SUN suggests to set q0 to about
10k. For benchmarkings as stated below, increasing the q is valid, too,
though not *that* high.

]could ever have 16000 connections waiting, if you have you're screwed
]anyway. For the FD's you're way over the top too, if the thing has need
]for over 4000 FD's there's something wrong.

I agree to this. Even while benchmarking there was no need for me for more
than 4096 FDs. 'Upgrading to Linux' is not always an option. Setting FDs
too high is said to have a negative impact on performance, too, e.g. 4096
should suffice.

]> tcp_time_wait_interval 60000
]> tcp_conn_req_max_q 16384
]> tcp_conn_req_max_q 16384
]> tcp_recv_hiwat_minmss 24
]>
]> FD is set to 12288.

There are a few more toggles and dials to fiddle with, especially when
benchmarking (so what systems (plural!) did you use to run your polygraph
suite from), but most won't do much difference. However, if you stuck to
the "how to polygraph" step-by-step as outlined by Alex,

        http://www.terena.nl/tech/d2-workshop/d2cache99/
        (at the very bottom)

you should have discovered during level 0 tests that you had to tune your
TCP buffers to 64k, otherwise you must have noticed that you weren't able
to achieve the netperf speed seen with FreeBSD. Until your level 0 tests
work out, all higher level results are bogus.

        http://www.rvs.uni-hannover.de/people/voeckler/tune/EN/tune.html
        (ok, special benchmark tuning still missing)

Le deagh dhùrachd,
Dipl.-Ing. Jens-S. Vöckler (voeckler@rvs.uni-hannover.de)
Institute for Computer Networks and Distributed Systems
University of Hanover, Germany; +49 511 762 4726
Received on Tue Jun 15 1999 - 06:04:56 MDT