I did the hdparm test again, as also before posting the first post.
The test is 3x repeated for accurate results.
Code:
# hdparm -Tt /dev/hdb
/dev/hdb:
Timing cached reads: 1848 MB in 2.00 seconds = 922.30 MB/sec
Timing buffered disk reads: 102 MB in 3.05 seconds = 33.47 MB/sec
# hdparm -Tt /dev/hdb
/dev/hdb:
Timing cached reads: 1872 MB in 2.00 seconds = 934.74 MB/sec
Timing buffered disk reads: 94 MB in 3.02 seconds = 31.14 MB/sec
# hdparm -Tt /dev/hdb
/dev/hdb:
Timing cached reads: 1856 MB in 2.00 seconds = 928.14 MB/sec
Timing buffered disk reads: 88 MB in 3.02 seconds = 29.18 MB/sec
The DMA is enabled:
Code:
# hdparm -d /dev/hdb
/dev/hdb:
using_dma = 1 (on)
Could it be the filesystem (ext3?)? The swap is sometimes even empty when the bottle neck occurs.
I have a Fedora Core 4 system, which is mostly compiled against i386, but the kernel is compiled against i686.
Is it a kernel problem, disk problem or a configuration problem?
When more processes use the disk, the scheduling algorithm for disk access goes nuts and the response times are pretty bad. Is there a way to set per process disk or i/o priority?
Answer to Emmanuel_uk:
The article link about iostat is
http://www.linuxquestions.org/questi...icle&artid=532
Another good utility for i/o monitoring is sar: it displays %iowait per daily intervals every 10 minutes.
But I still can't see per process i/o usage.