tunefs -a 32 /opt
dd if=/dev/zero of=/opt/q1 bs=1024k
iostat -xnz 5 2
extended device statistics
r/s w/s kr/s kw/s wait actv wsvc_t asvc_t %w %b device
0.0 82.1 0.0 18789.1 1.0 41.0 11.9 499.1 68 99 d10
0.0 83.3 0.0 19095.9 0.0 35.3 0.0 424.2 0 95 d11
0.0 81.5 0.0 18685.2 0.0 35.8 0.0 439.0 0 95 d12
0.0 84.9 0.0 19096.7 0.0 35.9 0.2 423.3 2 96 c0t0d0
0.0 82.9 0.0 18685.9 0.0 36.4 0.2 439.0 2 96 c0t1d0
tunefs -a 16 /opt
extended device statistics
r/s w/s kr/s kw/s wait actv wsvc_t asvc_t %w %b device
0.0 428.1 0.0 53325.8 0.6 61.5 1.4 143.8 59 98 d10
0.0 428.1 0.0 53325.8 0.0 61.1 0.0 142.8 0 97 d11
0.0 429.1 0.0 53453.8 0.0 58.8 0.0 137.1 0 96 d12
0.0 430.7 0.0 53327.2 0.0 61.7 0.0 143.3 0 100 c0t0d0
0.0 431.7 0.0 53455.2 0.0 59.4 0.0 137.6 0 99 c0t1d0
tunefs -a 8 /opt
extended device statistics
r/s w/s kr/s kw/s wait actv wsvc_t asvc_t %w %b device
0.0 835.5 0.0 52868.0 0.6 120.1 0.7 143.8 57 96 d10
0.0 835.5 0.0 52868.0 0.0 119.2 0.0 142.7 0 96 d11
0.0 835.3 0.0 52855.2 0.0 116.6 0.0 139.6 0 94 d12
0.0 838.1 0.0 52869.4 0.1 119.7 0.1 142.8 2 99 c0t0d0
0.0 837.9 0.0 52856.6 0.0 117.2 0.0 139.8 1 97 c0t1d0
tunefs -a 1 /opt
extended device statistics
r/s w/s kr/s kw/s wait actv wsvc_t asvc_t %w %b device
0.0 6019.8 0.0 48202.4 0.8 857.9 0.1 142.5 63 97 d10
0.0 6024.8 0.0 48242.4 0.0 840.3 0.0 139.5 0 96 d11
0.0 6004.2 0.0 48077.8 0.0 844.9 0.0 140.7 0 97 d12
0.0 6028.7 0.0 48244.4 127.0 222.2 21.1 36.9 80 98 c0t0d0
0.0 6008.2 0.0 48079.8 128.7 223.5 21.4 37.2 80 99 c0t1d0
update: actual testing on a production shows that SAS 10K disks comparing to SCSI 10K disks are 10-25% faster (mostyl writes, random IO).
Thursday, January 12, 2006
How fast are SAS disks?
I started to play with x4100 server. Its internal disks are SAS (Serial Attached SCSI). I've just made some simple test and... well, this disks can handle over 6000 IO/s while writing! It also looks like issuing large IOs is actually hurting performance - I'm not sure why.
Subscribe to:
Post Comments (Atom)
2 comments:
What options did you use to mount /opt? I suspect that 6000 I/O is from /dev/zero do memory cache not to disk itself. Give more details on server configuration, because it seems to be false result. There is no known single disk in the world to perform more than 200 IOPS (average) - so 6000 seems to be much to high value.
-C
The trick is that these are sequential IOs so these IOs are probably merged by disk itself. With random IOs you can't achive that much of course. The pojnt showing this was that you've got full bandwith to SCSI commands as well as for data using SAS disks so in some workloads these disks can be better than SCSI disks. With MySQL we have observed SAS 10K disks to be 10-25% faster than SCSI 10K disks.
Post a Comment