Hi, On Saturday, December 3, 2011 14:24 CET, Henning Brauer <lists-open...@bsws.de> wrote: > * Sebastian Reitenbach <sebas...@l00-bugdead-prods.de> [2011-12-03 01:12]: > > Thanks to your pointers on icb, I found the autosizing algorithm at the end > > of netinet/tcp_userreg.c. > > There I saw its comparing the actual value of fill level of the buffer, and > > how much got transferred against a > > maximum defined in sys/socketvar.h: > > > > #define SB_MAX (256*1024) > > > > With the default value of 256*1024, I got the maximum transfer rate of > > about 1.5MB/s. > > This speed there was fairly constant, when the maximum was reached. > > Then I doubled the value to 512*1024, and got constant transfer rates of > > about 2.7MB/s, which is about the same speed like I got with the Linux host. > > Then I again doubled the value to 1024*1024, and got speeds of about 5MB/s. > > This was not so constant anymore. Here I got spikes of up to 7MB/s, and in > > the middle it dropped down to 2-3 MB/s. > > Then again doubled the value to 2048*1024, and then I got transfer speeds > > of about 8MB/s. So about 2.5 times faster than with Linux as the server. > > Also here, the transfer rates are not constant over the download, they are > > varying from 3MB/s and going up to over 9MB/s, nearly to the maximum the > > network card in the server provides (100MBit). > > ok, so we're dealing with a high-latency high-bandwidth connection.
yep, as it states in the subject ;) > there bigger buffers and thus seqnr windows help a lot - due to the > latency. > > > Since my knowledge to the network stack is next to zero, I don't really > > know what side effects it would have, to raise the default to a higher > > value, and if it would acceptable to have those varying transfer speeds, > > instead of the constant rates with the lower values. > > Maybe a problem for machines with lower memory? > > "low memory" is a bit misleading, we're talking about pretty special > memory here - but yes, hundreds or thousands of sockets with 2MB buffers > in use each would be a problem. there is a backpressure mechanism in > the autosizing algorithm to (try to) avoid this; we might very well > want to increase SB_MAX a bit. Didn't had time today, I'll do some more testing, also even trying to find a gigabit card that I can put into the box, just to see if I can get even more speed... I'll then send a followup patch to tech@ thanks, Sebastian > > -- > Henning Brauer, h...@bsws.de, henn...@openbsd.org > BS Web Services, http://bsws.de, Full-Service ISP > Secure Hosting, Mail and DNS Services. Dedicated Servers, Root to Fully > Managed > Henning Brauer Consulting, http://henningbrauer.com/
