* XFS performance on LVM2 mirror
@ 2008-03-25 21:37 Scott Tanner
2008-03-29 11:17 ` Hannes Dorbath
2008-03-29 16:27 ` Michal Soltys
0 siblings, 2 replies; 3+ messages in thread
From: Scott Tanner @ 2008-03-25 21:37 UTC (permalink / raw)
To: xfs
Hello,
I've been doing some benchmarking (using bonnie++) of the XFS
filesystem and found a substantial performance drop in rewrite and
delete operations when using an LVM2 mirror. When using the Linux
Software Raid driver to perform the mirror, XFS performance is quite
good. I've tried a number of the performance tweaks from the mailing
list archives, as seen below. The only option that seemed to make a real
difference was LVM's --corelog, which only worked for 1 test then the
server crashed.
Are there any special tweaks for XFS on LVM2 mirrors? Recommendations on
my setup?
Thanks,
Scott
Here are some system specs and benchmark results:
Dell 1750 (32bit)
single Xeon 2.8GHz
1GB ram
Centos5 (2.6.18-53.1.14.el5)
Emulex lightpulse L982 HBA
2 x Jetstor Disk Arrays:
9 Disk Raid 5, 128K stripe
2 TB LU
16K readahead set on all disks (blockdev --setra 16384 /dev/sdx)
Adjusted max_sectors_kb - no improvement
Adjusted nr_requests - no improvement
XFS mounted with -o noatime,nodiratime,nobarrier,logbufs=8
All Bonnie++ tests using -s 2016M -n 2:5000000:1000000/64, displaying
average of 3 runs.
---------------------------------------------------------------------
Single 2TB disk (LU) no LVM no XFS options:
fs01,2G,40374,,186139,52,84250,24,43405,98,191468,14,745.0,0,,27,19,51,7,1389,20,26,20,18,1,1089,18
(S.Out Block= 186M/s |S.Out Rewrite= 84M/s |S.In Block=191M/s |Del =
1389)
Linux Software Raid Mirror, no XFS options:
fs01,2G,39516,,93718,25,75745,22,43253,98,193540,23,1475.8,1,,24,19,36,7,1184,18,24,19,17,2,969,12
(S.Out Block= 93M/s |Seq Out Rewrite= 75M/s |S.In Block=193M/s |Del =
1184)
LVM2 Mirror, no XFS options:
fs01,2G,33459,,93708,28,18752,5,42716,97,196157,20,1031.9,0,,25,21,33,5,180,2,26,22,18,1,226,1
(S.Out Block= 93M/s |S.Out Rewrite= 18M/s |S.In Block=196M/s |Del = 180)
LVM2 Mirror, XFS options logsize=64M sunit=256 swidth=512 blks
fs01,2G,34492,,93389,27,18381,5,43100,97,191331,21,1063.2,0,,23,20,29,4,178,1,24,19,17,1,191,1
(S.Out Block= 93M/s |S.Out Rewrite= 18M/s |S.In Block=191M/s |Del = 178)
LVM2 Mirror XFS options logsize=64M sunit=64 swidth=1024 blks
fs01,2G,36145,,92566,26,18430,5,43560,99,200523,21,1088.4,0,,23,20,34,5,211,2,25,20,18,1,172,1
(S.Out Block= 92M/s |S.Out Rewrite= 18M/s |S.In Block=200M/s |Del = 211)
LVM2 Mirror XFS options logsize=64M sunit=32 swidth=256 blks
fs01,2G,36922,,96298,26,18832,6,43623,97,199477,26,1187.2,3,,27,20,33,11,235,4,25,19,20,6,158,3
(S.Out Block= 96M/s |S.Out Rewrite= 18M/s |S.In Block=199M/s |Del = 235)
LVM2 Mirror --metadatasize=64k, XFS options logsize=64M sunit=32
swidth=256 blks
fs01,2G,34508,,96358,25,19132,6,43817,98,200040,27,1202.5,3,,29,22,33,10,234,4,29,23,19,6,195,4
(S.Out Block= 96M/s |S.Out Rewrite= 19M/s |S.In Block=200M/s |Del = 234)
LVM2 Mirror --metadatasize=64k, XFS options logsize=64M sunit=256
swidth=256 blks
fs01,2G,36367,,91923,24,18639,6,43732,97,199305,26,1230.0,3,,26,19,31,10,212,4,30,22,19,6,151,3
(S.Out Block= 91M/s |S.Out Rewrite= 18M/s |S.In Block=199M/s |Del = 212)
LVM2 Mirror --corelog, XFS options logsize=64M sunit=256 swidth=256 blks
fs01,2G,39387,,95708,30,46650,15,43273,98,199461,21,1037.3,0,,25,20,29,4,889,10,26,20,17,1,927,14
(S.Out Block= 95M/s |S.Out Rewrite= 46M/s |S.In Block=199M/s |Del = 889)
And for comparison, EXT3 over LVM2 mirror:
fs01,2G,33028,,117260,75,61222,38,36994,86,196760,45,885.5,2,,33,60,33,11,143,5,31,57,21,7,147,6
(S.Out Block=117M/s |S.Out Rewrite= 61M/s |S.In Block=196M/s |Del = 143)
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: XFS performance on LVM2 mirror
2008-03-25 21:37 XFS performance on LVM2 mirror Scott Tanner
@ 2008-03-29 11:17 ` Hannes Dorbath
2008-03-29 16:27 ` Michal Soltys
1 sibling, 0 replies; 3+ messages in thread
From: Hannes Dorbath @ 2008-03-29 11:17 UTC (permalink / raw)
To: Scott Tanner; +Cc: xfs
Scott Tanner wrote:
> Are there any special tweaks for XFS on LVM2 mirrors?
LVM will always slow down things, it's an additional layer, there is no
way around that.
LVM does not do well with XFS stripe alignment, it causes sequential
writes to happen in a non-uniform, pumping way and additionally does not
support write barriers.
> XFS mounted with -o noatime,nodiratime,nobarrier,logbufs=8
Nodiratime is redundant. You might add logbsize=256k. Lazy-counters
require you to at least upgrade to 2.6.23.
I'm unsure if your stripe alignment is correct. Why don't you use the
su,sw options? Why do you run with only 64MB logsize?
> Recommendations on my setup?
Drop LVM altogether. Add ram to the box. Personally I'd use Solaris/ZFS
for that specific setup.
--
Best regards,
Hannes Dorbath
^ permalink raw reply [flat|nested] 3+ messages in thread
* Re: XFS performance on LVM2 mirror
2008-03-25 21:37 XFS performance on LVM2 mirror Scott Tanner
2008-03-29 11:17 ` Hannes Dorbath
@ 2008-03-29 16:27 ` Michal Soltys
1 sibling, 0 replies; 3+ messages in thread
From: Michal Soltys @ 2008-03-29 16:27 UTC (permalink / raw)
To: xfs
Scott Tanner wrote:
> Hello,
> I've been doing some benchmarking (using bonnie++) of the XFS
> filesystem and found a substantial performance drop in rewrite and
> delete operations when using an LVM2 mirror. When using the Linux
> Software Raid driver to perform the mirror, XFS performance is quite
> good. I've tried a number of the performance tweaks from the mailing
> list archives, as seen below. The only option that seemed to make a real
> difference was LVM's --corelog, which only worked for 1 test then the
> server crashed.
>
> Are there any special tweaks for XFS on LVM2 mirrors? Recommendations on
> my setup?
>
Recently - there have been plenty of topics on linux-raid mailing list
regarding this subject as well - worth checking them out.
Generally important things to consider - read-ahead on your lv volume
(or md device in case where xfs is directly on it - r-a of components
are not used then), alignment of logical volumes with respect to
underlying raid (especially as you use 128k chunk and by default volume
won't start at multiple of 128k), lv extent size, xfs su/sw parameters
which you will have to set manually, if you are on lvm on md raid.
^ permalink raw reply [flat|nested] 3+ messages in thread
end of thread, other threads:[~2008-03-29 16:44 UTC | newest]
Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2008-03-25 21:37 XFS performance on LVM2 mirror Scott Tanner
2008-03-29 11:17 ` Hannes Dorbath
2008-03-29 16:27 ` Michal Soltys
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox