[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: slapd and meta devices under Solaris 8

Michael Ströder wrote:
On 12:37:35 pm 2006-01-17 "Spicer, Kevin" <KevinS@bmrb.co.uk> wrote:
-----Original Message-----
From: "Michael Ströder"

I'm currently doing some tests with OpenLDAP on a system where the
real disk devices are bound together as RAID 1 meta devices. To me
slapd -q seems to be extremly slow on this systems (compared to
tests on my Linux i386 laptop).
After a while disks are 100% busy, CPU load is very low (<5%).

While slapadd running the disk load was 100%. Now it's 0. IMHO the Solaris LDAP client tools are not used at all.

Well, slapadd is a pretty I/O intensive operation, no matter how you slice it. The only meaningful comparison you could make is to measure it on the same Solaris system, but using a regular un-mirrored drive. Assuming your BDB cache is already as large as possible, there's not much else you can do. Note that with slapadd -q if you're using the tool-threads parameter, the indexing will be run in parallel. Once the BDB cache gets full, that will result in quite a lot more disk seek activity than otherwise, so this will definitely exercise your disks pretty intensely.

RAID1 is a mixed bag; you're fastest write is only as fast as the slowest drive in the array. It's possible to boost read rates; since the identical data is on all drives any large read request can be striped across the array, and most RAID 1 implementations will take advantage of this fact. But it has a cost: write rates will only be deterministic if all of the drives are identical and their seek positions are identical. If you're purely writing, these conditions will usually hold, but once you start striping reads across the array, that means each drive will have to seek a different distance than every other drive to accomplish a write - they are no longer writing in parallel lockstep. In the worst case, that means write rates are slowed down by a factor equal to the number of drives in the array.

That's why it would be interesting to compare your load performance to the same machine with a single drive...

 -- Howard Chu
 Chief Architect, Symas Corp.  http://www.symas.com
 Director, Highland Sun        http://highlandsun.com/hyc
 OpenLDAP Core Team            http://www.openldap.org/project/