From: Peter Wong <wpeter@us...> - 2001-10-01 22:42:31
In August I reported that there was a 40% performance gain on a
heavy database workload using Jens Axboe's zero-bounce highmem
I/O patch and my IPS patch running 2.4.5 on a 4-way machine.
This time I want to report the performance gain using the above two
patches and Jonathan Lahr's io_request_lock patch running 2.4.6 on
an 8-way machine.
With Jens' patch and my patch, the total elapsed time for processing
the queries was cut by 34%.
With the above two patches as the baseline, Jonathan's
io_request_lock patch cut the total elapsed time by an additional 36%.
The CPU utilization corresponding to the bounce patches was
25% (user), 63% (kernel) and 12% (idle).
With Jonathan's patch added, the CPU utilization became
40% (user), 29% (kernel) and 31% (idle).
We can see that there was a significant drop in kernel time and an
increase in idle time. In addition, I did not detect any system
problems using Jonathan's patch during a lengthy period of heavy
Jonathan's io_request_lock patch can be found at
Wai Yee Peter Wong
IBM Linux Technology Center, Performance Analysis
Office: (512) 838-9272, T/L 678-9272; Fax: (512) 838-4663