John Popplewell writes: As part of the process of fuelling the debate on Raid performance, I thought I would offer you the following practical experience of the performance of a system I have been looking at.

The system is running a large online transaction processing system at 600 message pairs per second and it is required to achieve more throughput. The first obvious step of adding further cpu power to the system didn't get any more throughput.

Analysis of one of the I/O subsystems [CLARiiON FC4700 with 10 physical 73 Gb disks configured in a Raid 1/0 configuration] showed that this was an extremely write intensive workload that was performing 2200 logical I/Os per second. Taking into account the percentage of the I/O's that performed updates and the fact that the discs are mirrored meant there was a requirement to destage from the cache approximately 3000 I/Os per second to the physical discs. The disk subsystem was unable to cope at this rate and the reason why more throughput couldn't go through the system was the fact that the database logs were also on the same disk subsystem.

The answer to more throughput on this system is to install additional physical discs on the CLARiiON disk subsystem.

Whilst accepting that the performance of any disk subsystem is a function of many items (number of spindles, cache size, disk spindle characteristics etc) I would suggest that as a rule of thumb the write rate per physical spindle should be kept below 150 I/Os per second to achieve "good" performance. I won't attempt to define "good performance".

The ROT of 150 write I/Os per second may need to be modified depending on the number of physical reads per second occurring.

Next UNIX Tip