Messages in this thread | | | From | Niels Provos <> | Date | Sat, 22 May 1999 16:10:36 -0400 | Subject | Re: hinting system for poll (was Re: /dev/poll vs. aio_) |
| |
In message <199905212222.PAA03854@pizda.davem.net>, "David S. Miller" writes: >Run 2.2.9, this tcp_tw_deschedule bogosity will go away. I just ported the hinting system over to 2.2.9 and things look nicer now.
I just did a very quick benchmark. The outputs of readprofile are probably biased in favor of the unmodified system because lacking file descriptors it skipped about 600,000 requests (see the difference in generic_copy_from_user).
httperf on local system 1,000,000 requests with rate 1200 r/s:
no hinting: min 0.0 avg 419.1 max 1104.6 stddev 168.2 (177 samples) hinting: min 829.8 avg 1197.3 max 1252.2 stddev 37.4 (167 samples)
And the readprofile output is:
non-hinting: 18132 sock_poll 453.3000 13874 inet_poll 315.3182 3933 synchronize_irq 196.6500 15200 tcp_poll 63.3333 18830 do_select 32.6910 1162 sock_close 19.3667 1279 tcp_write_space 15.2262 1146 synchronize_bh 15.0789 894 __generic_copy_from_user 14.9000 496 system_call 8.8571
hinting: 6080 synchronize_irq 304.0000 2090 __generic_copy_from_user 34.8333 34394 do_select 30.5996 1577 system_call 28.1607 1981 synchronize_bh 26.0658 140 ret_with_reschedule 8.7500 973 do_gettimeofday 8.1083 816 __wake_up 7.8462 183 signal_return 7.6250 241 sock_wfree 7.5312 [...] 126 inet_poll 2.2500 105 sock_poll 2.1875 417 tcp_poll 1.0322
Well, nontheless, I hope you find this interesting.
Greetings, Niels.
-- Niels Provos <provos@citi.umich.edu> finger provos@umich.edu for pgp info
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |