2.4.9-ac10 IDE access slows as uptime increases

2.4.9-ac10 IDE access slows as uptime increases

[George Garvey]

[-- Attachment #1: Type: text/plain, Size: 2956 bytes --]

   I've been noticing this for months, and thought I was crazy. But I
just verified it.
   I ran a program that's a GUI app/front-end to a data base, on the
local drives. It took seconds to access a record.
   Rebooted. Ran the same thing again. The program accessed records
almost instantly.
   I get faster response time from the 2.2.20-pre2 systems over 100MBs
NFS than I do on the 2.4 system after it has been running for a few
hours. NFS is slow enough, usually use ssh and run the program on the DB
system. (Only use a local copy when testing.)
   Use the system mostly to compile, debug and test.

processor	: 0
vendor_id	: GenuineIntel
cpu family	: 6
model		: 8
model name	: Pentium III (Coppermine)
stepping	: 10
cpu MHz		: 999.743
cache size	: 256 KB
fdiv_bug	: no
hlt_bug		: no
f00f_bug	: no
coma_bug	: no
fpu		: yes
fpu_exception	: yes
cpuid level	: 2
wp		: yes
flags		: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 mmx fxsr sse
bogomips	: 1992.29

        total:    used:    free:  shared: buffers:  cached:
Mem:  1053401088 1002246144 51154944 10092544 51945472 798220288
Swap: 2147467264     4096 2147463168
MemTotal:      1028712 kB
MemFree:         49956 kB
MemShared:        9856 kB
Buffers:         50728 kB
Cached:         779508 kB
SwapCached:          4 kB
Active:         505116 kB
Inact_dirty:     76400 kB
Inact_clean:    258580 kB
Inact_target:      352 kB
HighTotal:      130992 kB
HighFree:         7256 kB
LowTotal:       897720 kB
LowFree:         42700 kB
SwapTotal:     2097136 kB
SwapFree:      2097132 kB

Personalities : [raid1] 
read_ahead 1024 sectors
md0 : active raid1 ide/host0/bus1/target1/lun0/part2[1] ide/host0/bus0/target0/lun0/part2[0]
      6291840 blocks [2/2] [UU]
md1 : active raid1 ide/host0/bus1/target1/lun0/part3[1] ide/host0/bus0/target0/lun0/part3[0]
      51661952 blocks [2/2] [UU]
unused devices: <none>

  8:32pm  up  1:42, 23 users,  load average: 2.87, 3.07, 2.77
All those users are me, logged into different xterms, most of which are
not being used during the uptime.

   XFree86 4.0.3. Using nVidia drivers. I really don't remember if the
problem was there when I was using an Athlon similarly configured except
for a Voodoo3 and 3/4 the RAM. Had so many problems with the Athlon, I
gave it to a co-worker who doesn't use their box as heavily as I do.
They haven't had any problems with the Athlon.
   I seem to remember this as a problem on the Athlon, too, but won't
swear to it because there were too many other things going on and this
is the first time I actually checked to see if the slowdown was real.
   Also have VMware modules loaded all the time. Don't always use it
every day though, and still see the slowdown. The system is rebooted
each morning right now, most of the time (unless I just forget). The
slowdown doesn't just affect the data base program. Its just really
noticeable there.
   Config attached. Any ideas how to check this out further?

[-- Attachment #2: config.gz --]
[-- Type: application/x-gunzip, Size: 4748 bytes --]

Re: 2.4.9-ac10 IDE access slows as uptime increases

[Alan Cox]

>    I've been noticing this for months, and thought I was crazy. But I
> just verified it.
>    I ran a program that's a GUI app/front-end to a data base, on the
> local drives. It took seconds to access a record.

Is the data base doing I/O directly to a block device and not using
O_DIRECT for one question

Second question is what is in your IDE logs. The IDE layer will change
down speeds when it hits a repeated problem (eg a DMA timeout) so if
need be will switch back to PIO or to MWDMA.

Re: 2.4.9-ac10 IDE access slows as uptime increases

[Phil Blecker]

   DB uses regular file system (ext3) and mmap.
   Nothing in logs from IDE. Only thing there unknown to me from kernel:
	cmpci: dma timed out??
That's from the sound driver, apparently. No idea what it means.
   By the way, would hdparm -i report the change? If so, they're all
still UDMA.
   I'm going to try removing the IDE to SCSI stuff that I installed to
try and get a USB CDRW drive working. Never did, and have not used any
time to pursue it (always timed out on access, now not even in devfs for
some reason). No idea if that's relevant, either, but it looks like
something I botched.

   The last message was so long that this may have been missed: it's not
just the DB that shows the slowdown. It's general. It's just really
obvious with the DB due to visual feedback. Who would notice if gcc-3
took an extra 20 minutes to compile something ;( (I haven't really timed
it. I noticed gcc gets slower, but not timed it to see how much, so its
subjective.)

   This may be irrelevant, but here's meminfo this morning:
        total:    used:    free:  shared: buffers:  cached:
Mem:  1053401088 887148544 166252544    65536 242728960 552976384
Swap: 2147467264 74866688 2072600576
MemTotal:      1028712 kB
MemFree:        162356 kB
MemShared:          64 kB
Buffers:        237040 kB
Cached:         478604 kB
SwapCached:      61412 kB
Active:         421960 kB
Inact_dirty:    352208 kB
Inact_clean:      2952 kB
Inact_target:      260 kB
HighTotal:      130992 kB
HighFree:         2036 kB
LowTotal:       897720 kB
LowFree:        160320 kB
SwapTotal:     2097136 kB
SwapFree:      2024024 kB

On Sat, Sep 29, 2001 at 03:33:17PM +0100, Alan Cox wrote:
> >    I ran a program that's a GUI app/front-end to a data base, on the
> > local drives. It took seconds to access a record.
> 
> Is the data base doing I/O directly to a block device and not using
> O_DIRECT for one question
> 
> Second question is what is in your IDE logs. The IDE layer will change
> down speeds when it hits a repeated problem (eg a DMA timeout) so if
> need be will switch back to PIO or to MWDMA.

Re: 2.4.9-ac10 IDE access slows as uptime increases

[Steve Maughan]

* Alan Cox (alan@lxorguk.ukuu.org.uk) wrote:
> Second question is what is in your IDE logs. The IDE layer will change
> down speeds when it hits a repeated problem (eg a DMA timeout) so if
> need be will switch back to PIO or to MWDMA.

I have what sounds like a similar problem - assuming it is a problem
with the IDE layer dropping the speed - except mine is with my CD or
DVD drive. They run a lot slower under the 2.4.x kernels than under
2.2.19 (looking at 2-3megs/s under 2.4.x, 5-7megs/s under 2.2.19).

Also when using 2.4.x, the kernel has a habit of disabling DMA and
performance plummets. Yet it is fine with 2.2.19.

Is this a related issue? Is there a fix for this?

Steve Maughan

Re: 2.4.9-ac10 IDE access slows as uptime increases

[Alan Cox]

> I have what sounds like a similar problem - assuming it is a problem
> with the IDE layer dropping the speed - except mine is with my CD or
> DVD drive. They run a lot slower under the 2.4.x kernels than under
> 2.2.19 (looking at 2-3megs/s under 2.4.x, 5-7megs/s under 2.2.19).
> 
> Also when using 2.4.x, the kernel has a habit of disabling DMA and
> performance plummets. Yet it is fine with 2.2.19.
> 
> Is this a related issue? Is there a fix for this?

That sounds different, Talk to Andre I guess

Re: 2.4.9-ac10 IDE access slows as uptime increases

[George Garvey]

On Sat, Sep 29, 2001 at 08:53:23AM -0700, Phil Blecker wrote:
>    I'm going to try removing the IDE to SCSI stuff that I installed to
> try and get a USB CDRW drive working.

   That seems to have worked. Haven't used the system as heavily today
as usual, but the slowdown didn't happen. Makes no sense to me, but what
do I know?