KeyWords: (see the KeyWords page for guidelines on what to put here.)

Performance measurements from a single Thor

To be able to make long-term calculations about bandwidth requirements in the Phoenix Cluster we need an accurate and round number to work with, when talking about how much data bandwidth can a Thor give.

Disk layout

  • There are 5 Raid6 arrays made of 7+2 disks, and three spare disks
  • Those five arrays are joined together in a volume group, then a logical volume, and then a xfs filesystem
  • All with default values.

Performance

The disk performance to be around 700 MB/s using Bonnie++

Then, we tried to measure how much the clients can get out of it. We use dccp transfers to /dev/null

  • With 2 wn, 15 or 20 transfers per WN, we get 450 MB/s
  • Above that, even with 90 transfers per WN, we get 500 MB/s.
  • One single WN with 45 transfers gets only 310 MB/s
  • Putting three WNs doesn't increase the 500 MB/s boundary
  • The write speed for one WN and 45 transfers at the same time is 70 MB/s

All in all, we can determine that each Thor can serve up to 500 MB/s read speed to the clients.

Another disk layout - with raid0

We realised that, from a single transfer, you only access one of the five raid6s. That's because creating a volume group is not the same as creating a stripped raid0 (looks like it creates a jbod). So, we tried making a raid0 on top of all raid6s. The rest is the same as before

Performance

  • Write speed, from 2 WNs and 23 transfers each, we get 270 MB/s write speed peaks.

Readers' comments

 


This topic: LCGTier2 > WebHome > PhoenixClusterBlog > PhoenixBlog20100831ThorPerformance
Topic revision: r1 - 2010-08-31 - PabloFernandez
 
This site is powered by the TWiki collaboration platform Powered by Perl This site is powered by the TWiki collaboration platformCopyright © 2008-2024 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback