NFS root, file server, all on sparc...

[shimitar]

Hallo....
i am using an Ultra5 to export the root filesystem for my intel base system. The goal was to remove hdd's from my PC and drastically reduce rumorosity.

I had been completely successfull on this regard... my pc boots and works perfectly without dics and even the performace is "nice".

More details, in case someone is interested:
1) i am booting the PC from floppy using ISOLINUX to give the right parameters to the ethernet (i do not want a tftp server nor dhcp or bootp)

2) The sparc box runs Gentoo and exports the root filesystem. BEWARE! disable any TCP based nfs... or your PC will hang whenever the disks on the sparc glitches (will explain more later)

3) connect the machines using 100mb/sec ethernet.

4) make whure you have at least 128mb of ram on the ultra5... and 512mb/1gb on your PC, you do jnot want to SWAP over ethernet!

5) raw performace is not bad, i achieve about 8/9mb/sec but bunch of small files are a slow business...

6) AVOID RAID! on sparc, raid0 chunk sizxe is 8kb, incompatible with x86 raid0! I am not really clear on this, but if you do not want to loose all your data on the raid just because your sparc is died... avoid software RAID, or you will not be able to mount the raid on a x86 machine.

7) You cannot get much performance from the onboard IDE controller ojn the ultra5... enable MDMA (hdparm -d1 -X66 -u1 -c1) but the good news is: NFS always does block transfer even on small accesses so on the client you always get the maximum performace.

gigabit ethernet is useless (9mb/sec througput out of 12 achievable on UDP)

I got some troubles....

at first, i checked "nfs over tcp" on the sparc box, and suffered from many system hangs on the PC client.

I discovered one of the disks, the boot disk, was corrupted and caused a few IDE0 channel resets affecting the second driver, the one with my exported root. Apparently, the NFS over tcp subsystem was not able to recover and even if after a few seconds the sparc system was usable again with mo flaws, the PC client was hang forever (= about 30mn).

So i unchecked nfs over tcp and now i am running nfs over UDP, i do not have any more tourbles... even after an IDE0 channel reset the PC client is able to resume its operations in a couple of seconds.

Finally, i had to replace the bad drive... but it did not affected the sparc system because only the kernel and silo.conf was on it and where unaffected from the disk damage.

Just to share with you all guys....

byebye
_________________
Willy Gardiol
willy@gardiol.org

[bazik]

[shimitar]

Ok, you can set chunk size, but i tried to move a RAID0 created on an intel based machine on the sparc (4kb chunck) and discovered i cannot use 4kb size on sparc!

So i created the raid0 under sparc with chunk 8kb and moved on the intel and discovered the intell will not be able to recognize it!

Maybe some issues with endianess? i am a bit confused, but resemble sparc has different "gender" about this issue.
_________________
Willy Gardiol
willy@gardiol.org

[hageshii]

I totally second that point about disabling TCP-based NFS.

I had been running NFSroot over TCP on my 10/100 setup for a couple years with no problems. But when I upgraded to gigabit, my workstation started hanging during disk-intensive things (emerging this-or-that), usually at consistent times.

I suspected the network link and removed the 'tcp' option from my kernel command-line, and things straightened up.