Tobias Rapp wrote: > Hello Arno, > >>> By default winsock's send buffer size is 8 kb which is much bigger >>> than TFtpCli's constant BLOCK_SIZE of 1460 bytes. So one condition >>> described in the article is not true since a single send call in >>> TFtpCli does not fill the whole underlying winsock send buffer. >>> I realy wonder why disabling Nagle doesn't help? >> >> Same applies to WSocket's internal send buffer that has a block size >> of 1460 as well. So if you haven't increased BufSize of TWSocket >> a single send call does never fill the whole underlying winsock send >> buffer. > > That's exactly the thing I'm doing at the moment: I increase the > BufSize of TWSocket from 1460 bytes to 1MB. The throughput increases > from 1.2MB/s to more than 7MB/s.
Ahh, I was thinking that you increased winsock's send buffer size. With a BufSize > winsock send buffer you risc to fullfill the error conditions described in the article... I would play with sligtly increased SocketSndBufSize and a BufSize set to something below like 16/8 and 32/16 kb. TWSocket's property BufSize actually is the block size and the amount of data copied into the winsock buffer in a single (winsock) send call. Increasing SocketSndBufSize however also decreases the maximum posible sockets in the system (no problem if you don't need hundreds of sockets at the same time). > > As far as I understand Nagle=off only would help me if I would like to > send small packets (< MCU) at a very high rate. But that is not the > case: I want to deliver large files (> 3GB). > > When I look at the IP traffic with Wireshark I see something like this > in the case where TWSocket.BufSize=1460 : > > Client->Server 1448 bytes > Client->Server 12 bytes > Server->Client ACK > Client->Server 1448 bytes > Client->Server 12 bytes > Server->Client ACK > .... > > And in the case where TWSocket.BufSize=1MB : > > Client->Server 1448 bytes > Client->Server 1448 bytes > Client->Server 1448 bytes > // previous line repeated 10-30 times > Server->Client ACK > Client->Server 1448 bytes > Client->Server 1448 bytes > Client->Server 1448 bytes > // previous line repeated 10-30 times > Server->Client ACK > .... > >>>> So shall I explicitly add a call setsockopt() that increases the >>>> socket send buffer dramatically (>1MB)? >>> Use WSocket option SocketSndBufSize instead, it will the call >>> setsockopt() internally. I think that 1 MB is much too big, since >>> this memory is allocated from kernel memory which is not an endless >>> resource. > > SocketSndBufSize property is read-only, at least in my version of > WSocket.pas (V5.25). > > /Tobias > > > -- > NOA Audio Solutions Vertriebsges.m.b.H. Tel: +43-1-5452700 > Johannagasse 42/4 Fax: +43-1-545270014 > A - 1050 Wien Www: http://www.noa- > audio.com -- To unsubscribe or change your settings for TWSocket mailing list please goto http://www.elists.org/mailman/listinfo/twsocket Visit our website at http://www.overbyte.be
