Per several correspondents, setting the maximum possible TCP buffer is a good start, but the application may or may not take advantage of this. In Samba's case, one can tell Samba to grab a larger buffer: /etc/samba/smb.conf [global] socket options = SO_RCVBUF=65536 And in fact, I can see the Solaris box advertising a TCP Window size of 64,240 bytes now, when I examine packet traces of file copies over SMB. A more complete socket options line might look as follows: [global] socket options = TCP_NODELAY SO_SNDBUF=65536 SO_RCVBUF=65536 Thank you for your assistance, --sk On 8/8/2010 5:09 PM, Stuart Kendrick wrote: > Hi folks, > > I'm setting TCP Receive Window size (along with other parameters) in a > custom /etc/init.d script: > > /usr/sbin/ndd -set /dev/tcp tcp_max_buf 4096000 > /usr/sbin/ndd -set /dev/tcp tcp_xmit_hiwat 4096000 > /usr/sbin/ndd -set /dev/tcp tcp_recv_hiwat 4096000 > /usr/sbin/ndd -set /dev/tcp tcp_wscale_always 1 > /usr/sbin/ndd -set /dev/tcp tcp_tstamp_if_wscale 1 > /usr/sbin/ndd -set /dev/tcp tcp_cwnd_max 4096000 > > Seems successful. > > > /usr/sbin/ndd -get /dev/tcp tcp_recv_hiwat > 4096000 > > > > But, for grins, I reboot anyway. > > > /usr/sbin/ndd -get /dev/tcp tcp_recv_hiwat > 4096000 > > > > However, when I negotiate a TCP session with the box, I see TCP Receive > Window sizes ranging around 48-49K, as seen in a packet trace, same as I > what I saw prior to fiddling with tcp_recv_hiwat > > https://vishnu.fhcrc.org/fred/telnet-to-fred5.pcap > https://vishnu.fhcrc.org/fred/shiva-to-fred5-smb-negotiate.pcap > > > I would expect the box to advertise a TCP Receive Window size of > 4096000, not 49232 and 49202 (telnet/TCP negotiation) or 49680 and 49532 > and 48300 (SMB/TCP negotiation). And for it to use RFC1323 window > scaling to do so. > > Any ideas why not? > > > uname -a > SunOS fred5 5.10 Generic_142900-05 sun4v sparc SUNW,Sun-Fire-T200 > > > > --sk > > Stuart Kendrick > FHCRC > > P.S. The box is slow at receiving files ... I can see TCP Receive Window > size dropping to 0 regularly ... > http://vishnu.fhcrc.org/fred/comparing-tcp-window-size-io-graphs.pdf > This might, of course, stem from slow storage or some other cause, such > that the box cannot drain the TCP buffer fast enough. But then again, > this is a 1GigE environment, with fast clients, and by my BDP > calculation, I need at least a 64K window size in order to keep up with > them, possibly closer to 128K. Ergo my efforts to fiddle with window size. _______________________________________________ sunmanagers mailing list sunmanagers@sunmanagers.org http://www.sunmanagers.org/mailman/listinfo/sunmanagersReceived on Mon Aug 9 19:00:15 2010
This archive was generated by hypermail 2.1.8 : Thu Mar 03 2016 - 06:44:17 EST