Automatic TCP send and receive socket buffer sizing

[Andre Oppermann]

王晓东 wrote:
> I have applied the patch of automatic send buffer sizing.
> My log file is:
> Dec 13 21:54:25 FreeBSD kernel: tcp_output: 202.119.117.246 inc sockbuf, old
> 33304, new 41496, sb_cc 30328, snd_wnd 65160, sendwnd 20308
> Dec 13 21:54:25 FreeBSD kernel: tcp_output: 202.119.117.246 inc sockbuf, old
> 41496, new 49688, sb_cc 39680, snd_wnd 66608, sendwnd 20272
> Dec 13 21:54:25 FreeBSD kernel: tcp_output: 202.119.117.246 inc sockbuf, old
> 49688, new 57880, sb_cc 46840, snd_wnd 65160, sendwnd 24616
> Dec 13 21:54:25 FreeBSD kernel: tcp_output: 202.119.117.246 inc sockbuf, old
> 57880, new 66072, sb_cc 54536, snd_wnd 65160, sendwnd 30408
> Dec 13 21:54:25 FreeBSD kernel: tcp_output: 202.119.117.246 inc sockbuf, old
> 66072, new 74264, sb_cc 63848, snd_wnd 65160, sendwnd 33304
> Dec 13 21:54:25 FreeBSD kernel: tcp_output: 202.119.117.246 inc sockbuf, old
> 74264, new 82456, sb_cc 71544, snd_wnd 65160, sendwnd 39096
> Dec 13 21:54:25 FreeBSD kernel: tcp_output: 202.119.117.246 inc sockbuf, old
> 82456, new 90648, sb_cc 73944, snd_wnd 66608, sendwnd 43440
> Dec 13 21:55:58 FreeBSD kernel: tcp_output: 202.119.117.245 inc sockbuf, old
> 33304, new 41496, sb_cc 29818, snd_wnd 64128, sendwnd 15461
> Dec 13 21:58:13 FreeBSD kernel: tcp_output: 202.119.117.45 inc sockbuf, old
> 33580, new 41772, sb_cc 31672, snd_wnd 65535, sendwnd 17520
> Dec 13 21:58:13 FreeBSD kernel: tcp_output: 202.119.117.45 inc sockbuf, old
> 41772, new 49964, sb_cc 37592, snd_wnd 65535, sendwnd 21900
> Dec 13 21:58:13 FreeBSD kernel: tcp_output: 202.119.117.45 inc sockbuf, old
> 49964, new 58156, sb_cc 44324, snd_wnd 65535, sendwnd 23360
> Dec 13 21:58:13 FreeBSD kernel: tcp_output: 202.119.117.45 inc sockbuf, old
> 58156, new 66348, sb_cc 53408, snd_wnd 65535, sendwnd 27740
> Dec 13 21:58:13 FreeBSD kernel: tcp_output: 202.119.117.45 inc sockbuf, old
> 66348, new 74540, sb_cc 62492, snd_wnd 65535, sendwnd 32120
> Dec 13 21:58:13 FreeBSD kernel: tcp_output: 202.119.117.45 inc sockbuf, old
> 74540, new 82732, sb_cc 70116, snd_wnd 65535, sendwnd 37960
> Dec 13 21:59:29 FreeBSD kernel: tcp_output: 202.119.117.245 inc sockbuf, old
> 41496, new 49688, sb_cc 36934, snd_wnd 64128, sendwnd 19101
> Dec 13 21:59:29 FreeBSD kernel: tcp_output: 202.119.117.245 inc sockbuf, old
> 49688, new 57880, sb_cc 44222, snd_wnd 69888, sendwnd 22976
> 
> And 117.45 is XP, 117.245 is Debian, 117.246 is FreeBSD-6-Stable.
> 
> How to analyze the performance? I need a help.

The log lines prove that the automatic send buffer sizing was
working.  On local high speed networks with very low RTT (<1ms)
the effect is almost not noticeable.  On links with higher
bandwidth and larger RTT (>10ms Internet and Corporate WAN)
there can be a very large difference.  You can perform a test
by transferring the same large file twice, once with auto sizing
enabled and once without.  A good target is the European half
of FTP.FreeBSD.ORG [62.243.72.50] as it has a high bandwidth
connection and the send buffer auto scaling patch installed.

-- 
Andre