Re: amdump performance problem
2007-03-07 13:36:50
On Wed, 7 Mar 2007 at 11:45am, Kenneth Berry wrote
The computer mbthome is both the amanda server and client, it is a Dell
PE2650 dual Xeon processor, 2GB Ram. Tape unit is LTO3 on a dedicated
SCSI controller. Internal HDD's are four SCSI ultra320 10K configured
as hardware RAID5. On top of this 100GB RAID5 disk is LVM Volume00
allocated as shown below. One additional internal SCSI ultra320 10K
drive of 300GB capacity is allocated to LVM Volume02 and usage can be
seen below. All the internal HDD share a common percraid controller. I
have an external RAID5 set of drives on an third SCSI controller which
LVM Volume01 resides. It should not be a factor currently it is unused.
Over the past weekend I migrated all the /home data from the external
set of drives to the single large internal drive, thinking the external
drives were the problem. But this reorganization made not difference.
Some quick things to look into:
1) Benchmark your various volumes with both bonnie++ (single thread
streaming) and tiobench (multithreaded). This will give you some baseline
numbers. You can also play with multiple instances of dd. I have no idea
how good those perc controllers are, as Dull tends to obfuscate them
compared to their LSI underpinnings.
2) LTO3 is *fast*. Reading from a 4 disk RAID5 should be able to keep it
streaming, but simultaneously writing to that same RAID5 is likely to be
*very* slow... Actually, looking at your tape write speeds in the
original mail, they're barely where they should be. LTO3 can throttle to
1/2 its native speed of 80MB/s before it has to resort to
drive/tape-lifetime-degrading stutter-stop-restart behavior. One thing
that significantly helps is increasing your blocksize from amanda's
standard 32KB -- I use 2MB. This requires recompiling amanda.
/dev/sdd1 141003764 36925740 96915448 28% /amanda
Erm, where does sdd fit in in terms of the LVM volumes?
Also, both of Jon's suggestions (spindle numbers and filesystem profile on
/home) are good ones to look into. Again, you can do your own
benchmarking outside of amanda to get a more controlled look at what's
going on.
--
Joshua Baker-LePain
Department of Biomedical Engineering
Duke University
|
|
|