dmlee@cheviot.uucp (Dongman Lee) (07/11/85)
I am going to increase the size of socket buffers on 4.2BSD. Before doing it, I would like to ask advice of wizards in netlands: i) what is the limit of the socket buffer size? ii) if I increase the size, does it affect the kernel operation? Thanx. Dongman Lee Computing Lab., University of Newcastle upon Tyne, UK ARPA: dmlee%cheviot.newcastle.ac.uk@ucl-cs.ARPA UUCP: <UK>!ukc!cheviot!dmlee -- Dongman Lee Computing Lab., University of Newcastle upon Tyne, UK ARPA: dmlee%cheviot.newcastle.ac.uk@ucl-cs.ARPA UUCP: <UK>!ukc!cheviot!dmlee
chris@umcp-cs.UUCP (Chris Torek) (07/12/85)
> I am going to increase the size of socket buffers on 4.2BSD. > Before doing it, I would like to ask advice of wizards in netlands: > i) what is the limit of the socket buffer size? > ii) if I increase the size, does it affect the kernel operation? There isn't really *a* "socket buffer size". There are a bunch of places that have various "recvspace" and "sendspace" variables (look for calls to soreserve()). These can be increased (and in fact many of them have been increased in 4.3) without affecting much other than the number of mbufs in use. I have heard somewhere that increasing tcp_sendspace and tcp_recvspace above about 12K causes a crash, but I haven't seen it happen (then again we haven't been running "plain" 4.2 for a long time). You might have problems if they are >16K, since things like windows are kept in shorts and signs of differences determine whether one number is "greater" or "less" than another... but 4K is lots anyway. -- In-Real-Life: Chris Torek, Univ of MD Comp Sci Dept (+1 301 454 4251) UUCP: seismo!umcp-cs!chris CSNet: chris@umcp-cs ARPA: chris@maryland
sylvain@lvbull.UUCP (Sylvain Langlois RCG-ARS) (07/15/85)
I don't think increasing socket buffer size is really necessary. I don't think there is a socket buffer size either !!. I've been looking at networking code of 4.2 BSD quite in detail ( I won't talk about AF_UNIX): and what is done about buffer size is only an initialization of so_hiwat at the beginning of new TCP connection. While implementing ISO Transport in 4.2BSD kernel, I increase this "size" by reserving something like 8K bytes for both queues. What I fear is that , if ever traffic is going to grow too much, all your mbufs pool is going to be eaten by socket buffers and some of your process are going to die because of ENOBUFS (except if your memory is big enough). I'm running with 30K bytes of mbufs and that even seems short to handle crisis situation ( like 4 users doing a transfer of /unix to their neighbours , some others playing with remote talk, well a fairly heavily loaded machine somehow!) Maybe another better idea would be increasing the data area of a mbuf? Any idea? -- Sylvain "Panic Trap" Langlois UUCP Address: (...!mcvax!vmucnam!lvbull!sylvain) Postal Address: BULL, PC 33/05, 68 route de Versailles, F-78430 Louveciennes, France.