Skip to contentSkip to conversation

Evaluation of three I/O schedulers

DISCUSSION Peter Wong 3 participants 3 messages Jun 30, 2003 – Jul 1, 2003

lore
Share & tools
lore

Peter Wong

author
#0 lore raw

We used 2.5.72+mm1 to evaluate three I/O schedulers, namely anticipatory, deadline and complete fair queueing under a very heavy database workload on an 8-way Pentium 4 machine. The workload is a decision support system doing mostly sequential I/O and each run takes about one hour. All three runs finished completely without encountering functional problems, and achieved similar performance level.

The 8-way machine has Pentium 4 2.0 GHz processors, 16 GB physical memory, 2MB L3 cache, 8 FC controllers with 80 disks. Hyperthreading was turned on for the three runs. The CPU utilization is similar for all three runs: 65% user, 7% system and 28% idle.

Regards,
Peter

Peter Wai Yee Wong
IBM Linux Technology Center Performance Team
email: [email protected]

#1 lore raw
Quoted 4 lines · Peter Wong

The 8-way machine has Pentium 4 2.0 GHz processors, 16 GB physical memory, 2MB L3 cache, 8 FC controllers with 80 disks. Hyperthreading was turned on for the three runs. The CPU utilization is similar for all three runs: 65% user, 7% system and 28% idle.

Could you give us an idea of how greatly utilized your 10 fiber controllers are? What about memory? Would a larger readahead be beneficial?

Signature · 2 lines
Dave Hansen
[email protected]
#2 lore raw
In reply to Peter Wong
Quoted 7 lines · Peter Wong

We used 2.5.72+mm1 to evaluate three I/O schedulers, namely anticipatory, deadline and complete fair queueing under a very heavy database workload on an 8-way Pentium 4 machine. The workload is a decision support system doing mostly sequential I/O and each run takes about one hour. All three runs finished completely without encountering functional problems, and achieved similar performance level.

Quoted 5 lines

The 8-way machine has Pentium 4 2.0 GHz processors, 16 GB physical memory, 2MB L3 cache, 8 FC controllers with 80 disks. Hyperthreading was turned on for the three runs. The CPU utilization is similar for all three runs: 65% user, 7% system and 28% idle.

Hi Peter,
How many block devices are being used at once in your tests? I would be interested to see profiles of AS and DL if possible. Thanks.

Nick