Date: Tue, 18 Jun 96 19:24:08 +0000 From: Andrew.Gordon@net-tel.co.uk To: cracauer@wavehh.hanse.de Cc: freebsd-hackers@FreeBSD.org Subject: Re(2): (SMB/Netware/NFS for DOS clients) (was BSD/OS ...) Message-ID: <"21933-960618192300-2ABD*/G=Andrew/S=Gordon/O=NET-TEL Computer Systems Ltd/PRMD=NET-TEL/ADMD=Gold 400/C=GB/"@MHS> In-Reply-To: <9606181038.AA28251@wavehh.hanse.de>
next in thread | previous in thread | raw e-mail | index | archive | help
> I've not been able to get acceptable performance out of samba. Not > more than 300-400 KB out of a 10Mbit-EThernet PCI 486 with 3com or WD > ISA ethernet card (or 200 KB/sec out of small Sparcs with SunOS). My > impression is that samba works fine, but is slow. Even PC-NFS on the > same machines reach a higher bandwidth, not to speak of a Windows NT > box serving NetBIOS (the on of our windows fraction fills 10 MB > ethernet easily). CPU time consumption is not the problem with samba, > the CPU is idle, the Ethernet unloaded, just the latency is too high > (as it seems to me). Are you using "-O TCP_NODELAY" on smbd? If not, this can give very high latency under some circumstances depending on the relative speeds of the machines involved. [cause is if the Nagle algorithm gets invoked on the final fragment of the transfer, causing that last fragment to be delayed until the TCP ack timer goes off. This is more likely to happen if the client is faster than the server, and also depends on interactions between the read size at the Windows machine, the buffer size used by Samba in copying from the file to the TCP socket, and the netw > If you think your samba servers are faster, could you please post some > benchmark results, reading and writeing a file in - say - 8 KB block > from a Win95 machine? Unfortunately, I don't have any sensible hardware here to test (nor DOS compilers), but here are some very coarse numbers on slow hardware. Performance on more sensible hardware in my office was much better, but I don't have numbers to hand. copy n:bigfile.dat NUL 25 sec = 400Kbyte/sec copy c:\bigfiledat n:bigfile2.dat 31 sec = 322Kbyte/sec Server is a 486DX2/66 with SMC Ultra ISA ethernet (FreeBSD2.1 + Samba), client is a 486/75 Toshiba portable, 3Com PCMCIA ethernet, Win95. I think these numbers are limited by the PCMCIA ethernet card. Using the same server, but another FreeBSD machine as client (486 with SMC PCI Ethernet this time), reading the same file gives these numbers with different protocols: NFS (cp bigfile.dat /tmp) 19 sec = 526 Kbyte/sec smbclient 14 sec = 666 Kbyte/sec (685 on screen) ftp 9 sec = 1111 Kbyte/sec (1.1Mb/s on screen) ["on screen" means as shown by the internal speed display of the program; all other times as measured by my wristwatch, so not very accurate]. This shows at least that the SMB protocol is not fatally speed-limited.
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?"21933-960618192300-2ABD*/G=Andrew/S=Gordon/O=NET-TEL Computer Systems Ltd/PRMD=NET-TEL/ADMD=Gold 400/C=GB/">