Hello,
I have the following configuration
SuSE Linux 7.3
Linux Kernel 2.4.x
4GB-Configuration
3 1/8GB RAM
216 GB HD (6*36)
Glibc 2.2.4
Informix 7.30 for Linux 2.2.x / glibc 2.1.3 (of course, for this
Kernel series and Library release)
400MB Shared Memory (non-resident)
IBM says, it works only with 2.2/glibc 2.1.3,
but we are forced by certain circumstances to use it under 2.4/glibc 2.2.x.
(the software we run doesn't work with higher versions of informix;
the number of users are too high to use it with lower memory)
At first, Informix works quite well, but after a while, especially
after some traffic on the computer (reading the harddisk by "find",
do a "update statistics" under informix or some exports, the system
starts thrashing.
I found out, that it starts thrashing with kswapd using 50% of processor
power and oninit using another 50%, when low memory runs short.
>From then, the system is very very slow.
The problem isn't so dramatic with kernel releases <=2.4.9, but with higher
releases, including 2.4.17, it is not tolerable.
2.4.17 does not swap to disk, but is still very slow, but kswapd is always
active (without swapping to disk!).
If i make the shared memory resident, the problem is worser than without.
There is always enough high memory free, before the problem occurs.
I could work around the problem by rebooting the computer automatically
in the morning (so that the problem doesn't occur in normal operation).
I don't know it, perhaps it is a silly question, but is it possible
to swap low memory different than high memory?
Or how can I resolve the problem?
Does anyone has an idea?
Greetings,
Martin Stolle
On Thu, 27 Dec 2001, Stolle, Martin (KIV) wrote:
> At first, Informix works quite well, but after a while, especially
> after some traffic on the computer (reading the harddisk by "find",
> do a "update statistics" under informix or some exports, the system
> starts thrashing.
>
> I found out, that it starts thrashing with kswapd using 50% of processor
> power and oninit using another 50%, when low memory runs short.
(sounds like much zone_normal unfreeable creating persistant imbalance)
> >From then, the system is very very slow.
> The problem isn't so dramatic with kernel releases <=2.4.9, but with higher
> releases, including 2.4.17, it is not tolerable.
> 2.4.17 does not swap to disk, but is still very slow, but kswapd is always
> active (without swapping to disk!).
What does /proc/slabinfo look like with 2.4.17? (zillion buffer heads?)
-Mike
On 2.4.9, buffer heads in /proc/slabinfo looks like
inode_cache 39855 40440 480 5055 5055 1 : 124 62
dentry_cache 13326 14490 128 483 483 1 : 252 126
filp 541 680 96 17 17 1 : 252 126
names_cache 3 3 4096 3 3 1 : 60 30
buffer_head 691683 977920 96 24445 24448 1 : 252 126
mm_struct 120 120 160 5 5 1 : 252 126
vm_area_struct 4299 4425 64 75 75 1 : 252 126
fs_cache 177 177 64 3 3 1 : 252 126
files_cache 117 117 416 13 13 1 : 124 62
signal_act 120 120 1312 40 40 1 : 60 30
buffer head grows always bigger and bigger.
btw,
I use raw partitions for storing the data.
Tomorrow you'll get the data for 2.4.17, I have to wait, because people
there
is a financial software on the database and the users want to enter the last
bookings
for the year 2001.
-----Original Message-----
From: Mike Galbraith
To: Stolle, Martin (KIV)
Cc: linux-kernel
Sent: 12/27/2001 8:42 AM
Subject: Re: Informix 7.3 and Linux Kernel 2.4
On Thu, 27 Dec 2001, Stolle, Martin (KIV) wrote:
> At first, Informix works quite well, but after a while, especially
> after some traffic on the computer (reading the harddisk by "find",
> do a "update statistics" under informix or some exports, the system
> starts thrashing.
>
> I found out, that it starts thrashing with kswapd using 50% of
processor
> power and oninit using another 50%, when low memory runs short.
(sounds like much zone_normal unfreeable creating persistant imbalance)
> >From then, the system is very very slow.
> The problem isn't so dramatic with kernel releases <=2.4.9, but with
higher
> releases, including 2.4.17, it is not tolerable.
> 2.4.17 does not swap to disk, but is still very slow, but kswapd is
always
> active (without swapping to disk!).
What does /proc/slabinfo look like with 2.4.17? (zillion buffer heads?)
-Mike