Seemingly odd disc i/o behaviour, need help to diagnose
dwhite at gumbysoft.com
Sat Jan 29 14:17:11 PST 2005
On Fri, 28 Jan 2005, Oren Baum wrote:
> Description of Server configuration and problems.
> FreeBSD 4.4-RELEASE
> Hard Drive (dual ATA disks, no RAID) on a Dell P4 PowerEDGE server
It bears mentioning again ... these releases are very old and have known
security issues. If this server is accessible to the public Internet its
probably been hacked several times over. An upgrade to at least
4.8-RELEASE+patches, apache 1.3.33, and the latest mysql-server-3.23 is
> We had many speed and timeout issues so we recompiled the kernel with
> maxusers=128 instead of the previous 32 and moved the hard drives into a new
> P4 2.8Ghz PowerEdge Case.
I'd think this is a PE1750 but those have SCSI disks, not IDE.
> atapci0: <Generic PCI ATA controller> port 0xffa0-0xffaf,0x374-0x377,0x170-
> 0x177,0x3f4-0x3f7,0x1f0-0x1f7 irq 11 at device 31.1 on pci0
> ata0: at 0x1f0 irq 14 on atapci0
> atapci1: <Generic PCI ATA controller> port 0xfea0-0xfeaf,0xfe30-0xfe33,0xfe20-
> 0xfe27,0xfe10-0xfe13,0xfe00-0xfe07 mem 0xdff3fc00-0xdff3ffff irq 5 at device
> 31.2 on pci0
> ata2: at 0xfe00 on atapci1
> ata3: at 0xfe20 on atapci1
Also as noted before this may be causing your disks to run at UDMA33 and
not at their full interface speed. The 'atacontrol mode ' command will
display the current settings (although you'll have to specify a bus, in
this case 0, 2, and 3).
> Server is configured to hold 2 medium sized MySQL DB's accessed through
> various perl and php scripts via websites on the server.
> Uptime: 78463 Threads: 18 Questions: 351685 Slow queries: 328 Opens: 2793
> Flush tables: 1 Open tables: 64 Queries per second avg: 4.482
> Each httpd process is of size 15MB to 20MB * about 15 processes
> Mysqld process is of size 32MB, resident 16MB
How much RAM does the machine have? A full dmesg output would be nice for
> If we run pine on a large mailbox or any other disk i/o intensive task,
> all other processes in motion seem to stall until the disk i/o is
> complete. This manifests itself in timeouts on webpages that require DB
> data, IMAP timeouts for other mail accounts, and even odd console/shell
I suspect this is simple I/O contention from one of two sources:
1. mysql keeps the disks loaded enough that when pine slurps in the big
file its able to out-contend the other processes.
2. pine bloats badly with large mailboxes, enough to cause other processes
to swap, which kills performance dead.
I'd suggest running 'iostat 1' in one window or VTY, 'top -s1' in another,
watch for a bit to get a baseline, then start up pine in another window
and watch the numbers change. In the top window, watch for the "Swap"
line to change with +In and +Out lines, and for the total use to go up. In
the iostat window watch how big MB/s gets, and the 'sy' column on the far
right, which is system CPU time as a percent.
If you're seeing the iostat MB/s go up to a level and stop, but swap stay
still, then its case #1. If you see iostat go up and Swap to start
increasing then its case #2. In case #2 the System CPU % should go up
radically (+10-20%) compared to baseline.
To fix case #1 you'll need to get faster storage, either by upgrading the
OS to make full use of the available ATA channels or by getting a more
capable storage subsystem (SCSI+disk array, for example). To fix case #2
you need to install more physical RAM.
ATA does not handle multiple transaction streams well since each bus can
only take one command at a time. The ata driver attempts to optimize the
order of these operations but compared to SCSI will not scale well in the
face of large quantities of diverse transactions. (Before people get up in
arms, I'll say that YMMV depending on workload, available cache, interface
There is, of course, option #3 which is to move your mail to another
Doug White | FreeBSD: The Power to Serve
dwhite at gumbysoft.com | www.FreeBSD.org
More information about the freebsd-stable