Raid0 slowdown from 2.4.19-rc1

Raid0 slowdown from 2.4.19-rc1

[Walt H]

Hello,

I've recently upgraded kernels from 2.4.19-rc1 to rc2 and on to final 
and have experienced a dramatic slowdown in raid0 performance since. 
Final 2.4.19 was patched with XFS and preempt and compiled using 
gcc-3.1.1 - I've tried switching compilers to 2.96 (mandrake 8.2), and 
cutting out preempt patches. First md1 array consists of two partitions 
from hda & hdc. hdparm for both drives looks fine by themselves:

[root@waltsathlon walt]# hdparm -t /dev/hda

/dev/hda:
  Timing buffered disk reads:  64 MB in  1.66 seconds = 38.55 MB/sec
/dev/hdc:
  Timing buffered disk reads:  64 MB in  1.65 seconds = 38.79 MB/sec

However, using them combined as raid0, md1:

/dev/md1:
  Timing buffered disk reads:  64 MB in  1.44 seconds = 44.44 MB/sec

In 2.4.18 and up through 2.4.19-rc1 I saw 66-70MB/sec from this array. 
Starting in rc2 it dropped to the mid 40's. I've also ran bonnie++ and 
confirmed several significant drops, however, not as bad as you would 
think considering the results of hdparm.

Hardware:
Soyo Dragon+ MB w/ Athlon 1800+, 512MB Ram
Drives connected via onboard Promise PDC20265 in ultra mode
2x40GB IBM Deskstar 60GXP (I know....)


Anything else you need, just ask. Please CC me in replies as I'm not a 
member of the list. Thanks,

-Walt

Re: Raid0 slowdown from 2.4.19-rc1

[Walt H]

Sorry, should have said more about raid arrays. Drives are partitioned 
as follows:

hda1, hdc1 = 4GB
hda2, hdc2 = Extended part - remainder of drive
hda5, hdc5 = 2GB = raid1, md0 /boot
hda6, hdc6 = ~15GB = raid0, md1 /usr
hda7, hdc7 = ~15GB = raid0, md2 /home
hda8, hdc8 = 1.5GB = raid0, /
hda9, hdc9 = remainder = swap

I agree, it seems as though you could see preempt lower performance, but 
again, I haven't seen that either. In fact, the 2.4.18 kernel I was 
using before was compiled with preempt also and showed ~68MB/Sec on md1,md2.

As for changes I may have made to .config? Nothing new. 2.4.19-rc1 
compiled with xfs and preempt worked well also. I tried looking for 
differences in raid drivers, but there were none to the raid0 driver. 
ide-pdc202xx.c contained many changes, but I'm not a kernel hacker and 
couldn't spot anything that might have affected this. Odd that it shows 
up even under hdparm. Interestingly, when testing with bonnie++, the 
overall sequential output was similar to the higher performing older 
kernels. However, creates, deletes, and rewrites were all down 
significantly.

-Walt


Mark Hahn wrote:
>>Final 2.4.19 was patched with XFS and preempt and compiled using 
> 
> 
> it's easy to imagine cases where preempt would produce lower performance,
> though I haven't seen any hard evidence of that.
> 
> 
>>cutting out preempt patches. First md1 array consists of two partitions 
>>from hda & hdc. hdparm for both drives looks fine by themselves:
> 
> 
> are they the first two partitions in hda/c?
> 
> 
>>/dev/hda:
>>  Timing buffered disk reads:  64 MB in  1.66 seconds = 38.55 MB/sec
>>/dev/hdc:
>>  Timing buffered disk reads:  64 MB in  1.65 seconds = 38.79 MB/sec
> 
> 
> such a disk will normally degrade to around half that performance
> in the tail of the disk.
> 
> 
>>/dev/md1:
>>  Timing buffered disk reads:  64 MB in  1.44 seconds = 44.44 MB/sec
>>
>>In 2.4.18 and up through 2.4.19-rc1 I saw 66-70MB/sec from this array. 
>>Starting in rc2 it dropped to the mid 40's. I've also ran bonnie++ and 
> 
> 
> nothing else changed?
>