Dai Ngo wrote: [...] >> I have just updated them to 1024000 (on both client and server). >> henrik at opensolaris:/# ptime cp /tmpfs/testfile_5g /nfs >> >> real 1:17.571581221 >> user 0.016147465 >> sys 13.271892825 >> >> That gives ~66 mb/s. >If I read your previous mail correctly, then you got 170 MB/s with >the previous TCP window setting (400000) and with 1024000 you >got 66 MB/s?
No - the obtained speed was dependend on the actual operation with 170 mb/s beeing the most I ever could get through over a single TCP connection. TCP window set to 400000 : A simple cp yields ~50 mb/s. A simple mv yields ~86 mb/s TCP window set to 1024000 : A simple cp yields ~82 mb/s A simple mv yields ~86 mb/s Read speeds are higher, ~220 mb/s on average. I am getting the feeling that this is not network related. FTP pushes 620 mb/s which is as fast as local write speed and iperf reports ~6.3 Gbit/s on average. Debugging nfsd has not been fruitfull as most operations seem to take place inside the kernel - unfortunatly my dtrace magic is not very strong. >> - the most I have been able to write via NFS is 170 mb/s. >Something is wrong with this picture. >I'd ask again, how did you make sure the new setting takes effect on both >client & server system? Client : henrik at opensolaris:~# ndd -get /dev/tcp tcp_recv_hiwat; ndd -get /dev/tcp tcp_xmit_hiwat 1024000 1024000 Server : henrik at naz01:~# ndd -get /dev/tcp tcp_recv_hiwat; ndd -get /dev/tcp tcp_xmit_hiwat 1024000 1024000 Iperf detects this as well : TCP window size: 1000 KByte (default) >-Dai > -- Med venlig hilsen / Best Regards Henrik Johansen