On Fri, Jul 19, 2002 at 11:40:31AM +0300, Jukka Marin wrote:
> [About NFS lockups that I (and some other ppl) have been seeing]
> 
> On Mon, Jul 15, 2002 at 10:18:25AM -0400, Nathan J. Williams wrote:
> > First thing to try: Crank down the NFS block size, with
> > -r=4096,-w=4096 in /etc/fstab. It seems that there is an unfortunate
> > interaction between our default NFS block size, network stack, and
> > slower network cards.
> 
> This seems to help.
> 
> > What network cards are the various machines using?
> 
> Server: NetBSD 1.5R, Athlon 900 MHz
> fxp0 at pci0 dev 15 function 0: Intel i82557 Ethernet, rev 8
> fxp0: interrupting at irq 10
> fxp0: Ethernet address 00:02:b3:1f:76:55, 10/100 Mb/s
> 
> Client1: NetBSD 1.6D, Athlon 900 MHz
> rtk0 at pci0 dev 15 function 0: RealTek 8139 10/100BaseTX
> rtk0: interrupting at irq 11
> rtk0: Ethernet address 00:10:a7:06:19:14
> 
> Client2: NetBSD 1.6_beta4, P233 MHz
> tlp0 at pci0 dev 10 function 0: DECchip 21041 Ethernet, pass 1.1
> tlp0: interrupting at irq 10
> tlp0: Ethernet address 00:e0:29:05:43:63
> tlp0: 10baseT, 10baseT-FDX, 10base2
> 
> Client3: NetBSD 1.6_beta4, P150 MHz
> wi0 at pcmcia0 function 0: IBM Corporation, IBM High Rate Wireless LAN PC
> Card, Version 01.01
> wi0: 802.11 address 00:02:2d:0d:c1:1c
> wi0: using Lucent Technologies, WaveLAN/IEEE
> wi0: Lucent Firmware: Station (6.4.1)
> 
> If Client1 is hooked up to a 10 Mbps hub, NFS locks up unless I use
> -r=4096,-w=4096  _or_  TCP mount.  If I hook the same machine directly
> to a 100 Mbps switch, NFS works with or without the options.
> 
> NFS read performance was tested with dd and a 40+ MB file and 64 kB block
> size.
> 
> host			no opts		-T		-r=4096,-w=4096
> Client1/10Mbps		lockup		ok (692 kB/s)	ok (1006 kB/s)
> Client1/100Mbps		ok (7933 kB/s)	ok (8906 kB/s)	ok (5471 kB/s)
> Client2/10Mbps		lockup		ok (815 kB/s)	ok (848 kB/s)
> Client3/11Mbps		lockup		ok (440 kB/s)	ok (342 kB/s)

Hum, maybe your switch doesn't have enouth memory to buffer the NFS
bursts with 32k reads (this will be something like 22 or 23 packets),
when the next hop isn't at 100Mbs ?

-- 
Manuel Bouyer <bouyer@antioche.eu.org>
--