On Tuesday 28 October 2003 07:50, Nick Piggin wrote:
cliff white wrote:
On Tue, 28 Oct 2003 05:52:45 +0800I'm working with Randy to fix this. Attached is what I have so far. See how
Michael Frank <mhf@xxxxxxxxxxxxx> wrote:
To my surprise 2.6 - which used to do better then 2.4 - does no longer handle these test that well.We see the same delta at OSDL. Try repeating your tests with 'elevator=deadline' to confirm.
Generally, IDE IO throughput is _very_ uneven and IO _stops_ at times with the
system cpu load very high (and the disk LED off).
IMHO the CPU scheduling is OK but the IO scheduling acts up here.
The test system is a 2.4GHz P4 with 512M RAM and a 55MB/s udma IDE harddisk.
The tests load the system to loadavg > 30. IO should be about 20MB/s on avg.
Enclosed are vmstat -1 logs for 2.6-test9-Vanilla, followed by 2.6-test8-Vanilla (-mm1 behaves similar), 2.4.22-Vanilla and 2.4.21+swsusp all compiled wo preempt.
IO on 2.6 stops now for seconds at a time. -test8 is worse than -test9
For example, on the 8-cpu platform:
STP id Kernel Name MaxJPM Change Options
281669 linux-2.6.0-test8 7014.42 0.0 281671 linux-2.6.0-test8 8294.94 +18.26% elevator=deadline
The -mm kernels don't show this big delta. We also do not see this delta on
smaller machines
you go with it.
This has been done without running a kernel compile, by $ ti-tests/ti stat ub17 ddw 4 5000
Seems to be more even on average but still drops IO too low and then gets overloaded.
By "too low" I mean io bo less than 10000.
By overloaded I mean io bo goes much above 40000. The disk can do maybe 60000.
When io bo is much above 40000, cpu scheduling is being impaired as indicated
by vmstat stopping output for a second or so...