mirror_region_size default: 4096 six times faster than 512?
Moving several terabytes of data, I was again struck by the very large difference in performance between copying data in small chunks for large chunks. After seeing larger block sizes make dd eight times faster, I tried increasing mirror_region_size in lvm.conf:
pvmove with default mirror_region_size of 512: ~4-6 MB/s pvmove with increased mirror_region_size of 4096: ~22-32 MB/s
Although this was somewhat unscientific testing, I'm quite curious what results others get on their hardware. The effect was most pronounced for me on a system with a 9550SX-12. I'll do some testing on a 8888ELP next week.
If 4096 is OFTEN several times faster than 512, would it make sense to consider canging the default?
Signature · 11 lines
Ray Morris [email protected] Strongbox - The next generation in site security: http://www.bettercgi.com/strongbox/ Throttlebox - Intelligent Bandwidth Control http://www.bettercgi.com/throttlebox/ Strongbox / Throttlebox affiliate program: http://www.bettercgi.com/affiliates/user/register.php
Quoted 2 lines · Ray Morris
If 4096 is OFTEN several times faster than 512, would it make sense to consider canging the default?
Hard to know, really.
What could be useful though would be some tuning scripts, that run tests like that on a given installation and suggest tweaks like that.
Alasdair
Quoted 4 lines · Ray Morris
If 4096 is OFTEN several times faster than 512, would it make sense to consider canging the default?
What disks are underneath this? One of the new 4k sector types?
Signature · 2 lines
Les Mikesell
[email protected]512 byte sector consumer SATA drives from Seagate and Hitachi. Drives attached to a 9560 RAID card exhibited more difference than drives in a different system attached to the motherboard directly. As I mentioned, I haven't done extensive or highly scientific testing, but in general, across various hardware using various software, copying chinks of a few MBs seems to always beat copying bytes or KB per chunk.
Signature · 1 line
Sent from my Android phone with K-9 Mail. Please excuse my brevity.
Les Mikesell <[email protected]> wrote:
On Tue, Feb 21, 2012 at 10:29 AM, Ray Morris <[email protected]> wrote:
If 4096 is OFTEN several times faster than 512, would it make sense to consider canging the default?
What disks are underneath this? One of the new 4k sector types?
Signature · 2 lines
Les Mikesell
[email protected]Quoted 6 lines · Ray Morris
512 byte sector consumer SATA drives from Seagate and Hitachi. Drives attached to a 9560 RAID card exhibited more difference than drives in a different system attached to the motherboard directly. As I mentioned, I haven't done extensive or highly scientific testing, but in general, across various hardware using various software, copying chinks of a few MBs seems to always beat copying bytes or KB per chunk.
What kind of RAID? Anything that involves parity is going to require small writes to read and merge the new data into blocks and rewrite the parity.
Signature · 2 lines
Les Mikesell
[email protected]Quoted 6 lines
copying chunks of a few MBs seems to always beat copying
Quoted 1 line
bytes or KB per chunk.
What kind of RAID? Anything that involves parity is going to require small writes to read and merge the new data into blocks and rewrite the parity.
At least mdadm RAID1 and RAID5. Hardware raid5 was slow with 512, but not yet tested with 4096 or higher. I'll test it in a few days after a migration completes.
Signature · 16 lines
Ray Morris [email protected] Strongbox - The next generation in site security: http://www.bettercgi.com/strongbox/ Throttlebox - Intelligent Bandwidth Control http://www.bettercgi.com/throttlebox/ Strongbox / Throttlebox affiliate program: http://www.bettercgi.com/affiliates/user/register.php On Fri, 24 Feb 2012 09:57:30 -0600
Quoted 15 lines · Les Mikesell
Quoted 7 lines · Ray Morris
512 byte sector consumer SATA drives from Seagate and Hitachi. Drives attached to a 9560 RAID card exhibited more difference than drives in a different system attached to the motherboard directly. As I mentioned, I haven't done extensive or highly scientific testing, but in general, across various hardware using various software, copying chinks of a few MBs seems to always beat copying bytes or KB per chunk.
What kind of RAID? Anything that involves parity is going to require small writes to read and merge the new data into blocks and rewrite the parity.