[nSLUG] 2.6 kernel performance

Peter Cordes peter at cordes.ca
Thu Jun 23 18:47:15 ADT 2005


On Tue, Jun 21, 2005 at 10:17:04AM -0300, Donald Teed wrote:
> My Debian home server is set up on a regular ext3
> partition and has a couple of software raid 5 partitions
> for data.  While I still have an ability to boot up
> the Gentoo raid 1 partition, I though it would be a good
> time to experiment with a 2.6 kernel from Debian.
> 
> This is a machine that has always run 2.4 kernel,
> dual PIII with BX chipset.  I've used a 2.6 kernel
> before, but never with software raid (md) and SMP.
> I'm not interested in tweaking it for desktop performance,
> but rather for disk I/O and network performance.
> 
> Last night I tried building a vanilla 2.6 kernel a few times,
> experimenting with I/O scheduler settings and such.
> There are two things that happen with 2.6.12 which
> seem strange.
> 
> One is the load factor reporting.  The first number in the series
> seems to "start" at 1 rather than 0.  When the machine is
> relatively idle with a 2.4 kernel, I see numbers like : 0.04, 0.03, 0.00
> When I run a 2.6 kernel, the first number is at least 1.
> The top report shows processors that are about 98% idle
> while the load number is like that.

 processes stuck in disk-sleep count toward load average.  On totally
different hardware (dual Opteron), I was having the acpi kernel thread
actually use 100% CPU when the UPS serial cable was plugged in.  Apparently
my cluster vendors had seen that before, and suggested turning off ACPI
(which solves that problem.)  The problem also didn't happen with later
kernel versions, IIRC.

 tuning tips:  RAID5 likes to write in big chunks, so it doesn't have to
read anything to calculate parity blocks.  3ware controllers have built-in
controllers with 128MB of buffer RAM, so their advice to crank up the
readahead might not be useful for software RAID5, but you might want to try.
(See my other email).  And give the anticipatory I/O scheduler a try;  It's
the default, and has worked well for me on software RAID1.  I haven't seen
any of the performance warnings you mentioned.

-- 
#define X(x,y) x##y
Peter Cordes ;  e-mail: X(peter at cor , des.ca)

"The gods confound the man who first found out how to distinguish the hours!
 Confound him, too, who in this place set up a sundial, to cut and hack
 my day so wretchedly into small pieces!" -- Plautus, 200 BC
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 351 bytes
Desc: Digital signature
URL: <http://nslug.ns.ca/mailman/private/nslug/attachments/20050623/4efcfc25/attachment-0002.pgp>


More information about the nSLUG mailing list