BUG: softraid,ext3 lockup

BUG: softraid,ext3 lockup

[elmer]

2.4.22 this time, on Xeon x335, / is raid1+ext3, journal size=128M

dd if=/dev/zero of=/tmp/jama bs=1M count=1000

on other softraid servers I have, this gets load to 3-4 , performance is
bad, the same behaviour is present, but tolerable.
Dual Xeon(exact same HW and conf with 2 cpus) now is about tolerable also

single  Xeon, with or without ht, still gets very stuck, load goes quickly
7-9, ls /directoryNotInCache takes 10 secs, cpu is idle, related processes
are in D state, as / is on same raid,everything sucks like hell..
additionally, something that might give a clue: when I mount the partition
with noatime, load gets quickly to 30...40..50
Elmer.

Re: BUG: softraid,ext3 lockup

[Marcelo Tosatti]

On Fri, 3 Oct 2003 elmer@linking.ee wrote:

> 
> 2.4.22 this time, on Xeon x335, / is raid1+ext3, journal size=128M
> 
> dd if=/dev/zero of=/tmp/jama bs=1M count=1000
> 
> on other softraid servers I have, this gets load to 3-4 , performance is
> bad, the same behaviour is present, but tolerable.
> Dual Xeon(exact same HW and conf with 2 cpus) now is about tolerable also
> 
> single  Xeon, with or without ht, still gets very stuck, load goes quickly
> 7-9, ls /directoryNotInCache takes 10 secs, cpu is idle, related processes
> are in D state, as / is on same raid,everything sucks like hell..
> additionally, something that might give a clue: when I mount the partition
> with noatime, load gets quickly to 30...40..50

Weird.

Which kernel does not show this behaviour? 

Thank you

Re: BUG: softraid,ext3 lockup

[Elmer]

all kernels I have seen are slugish on ide-raid large raw write,

However, now I got some performance test:

dualprocessor server has hda,hdb ibm-hitatchi
	7200 rpm IC35L060AVV207-0, 1860kb cache
uniprocessor  has hda,hdc(was hdb,no diff) western digital
	7200 rpm WDC WD400BB-23DEA0 2000kb cache


Also tested with hdparm -W 1

server      | writeback              |writetrough
------------|------------------------|-----------
dual + ibm: |24 sec,40MB/s,load 1-2  |90 sec 11MB/s,2-4
uni + wdc:  |36 sec,27MB/s,load 1-2  |160 sec 6MB/s,4-7..40


with harddisc writeback the problem does not manifest itself

the problem is present when I do this directly to raw partition /dev/hda2,
(while raid is running on same discs with very slim load and swap being
off, 1GB mem and lots free)

So there are two possible causes now:

1. ServerWorks CSB5 Chipset (rev 93) or driver plus hardiscs
2. Kernel itself, gets too high io performance and everything else is
stuck. So when IO troughput gets a bit higher, the bottleneck goes away.
When some other processes want a little time, they all get very stuck and
load goes to hell.


Elmer.