2.2.19pre6aa1 degraded performance for me...

2.2.19pre6aa1 degraded performance for me...

[Sasi Peter]

...compared to 2.2.18pre19.

I use the IDE patch for my CMD648 card, and also 0.90 RAID.

What I have now (2.2.19pre6aa1+ide-1221):
[root@iq /root]# hdparm /dev/hd[aceg]

/dev/hda:
 multcount    = 16 (on)
 I/O support  =  1 (32-bit)
 unmaskirq    =  1 (on)
 using_dma    =  1 (on)
 keepsettings =  1 (on)
 nowerr       =  0 (off)
 readonly     =  0 (off)
 readahead    = 16 (on)
 geometry     = 13240/16/63, sectors = 13345920, start = 0

/dev/hdc:
 multcount    = 16 (on)
 I/O support  =  1 (32-bit)
 unmaskirq    =  1 (on)
 using_dma    =  1 (on)
 keepsettings =  1 (on)
 nowerr       =  0 (off)
 readonly     =  0 (off)
 readahead    = 16 (on)
 geometry     = 35324/16/63, sectors = 35606592, start = 0

/dev/hde:
 multcount    = 16 (on)
 I/O support  =  1 (32-bit)
 unmaskirq    =  1 (on)
 using_dma    =  1 (on)
 keepsettings =  1 (on)
 nowerr       =  0 (off)
 readonly     =  0 (off)
 readahead    = 16 (on)
 geometry     = 39560/16/63, sectors = 39876480, start = 0

/dev/hdg:
 multcount    = 16 (on)
 I/O support  =  1 (32-bit)
 unmaskirq    =  1 (on)
 using_dma    =  1 (on)
 keepsettings =  1 (on)
 nowerr       =  0 (off)
 readonly     =  0 (off)
 readahead    = 16 (on)
 geometry     = 39560/16/63, sectors = 39876480, start = 0

[root@iq /root]# free -o
             total       used       free     shared    buffers     cached
Mem:        392808     389948       2860          0      15732     177004
Swap:       308432          0     308432
[root@iq /root]# cat /proc/swaps
Filename                        Type            Size    Used    Priority
/dev/hda2                       partition       102812  0       0
/dev/hde2                       partition       102812  0       0
/dev/hdg2                       partition       102808  0       0
[root@iq /root]# smbstatus|grep -E '
[[:digit:]]+.[[:digit:]]+.[[:digit:]]+.[[:digit:]]+'|cut -c
30-39|sort|uniq|wc -l
     41
[root@iq /root]# ifspeed 1
1.578717MB/s 0000000000000000000000>
1.758066MB/s 0000000000000000000000000>
1.526362MB/s 000000000000000000000>
1.530314MB/s 000000000000000000000>
1.659107MB/s 00000000000000000000000>
1.722287MB/s 000000000000000000000000>
1.542517MB/s 0000000000000000000000>
1.585877MB/s 0000000000000000000000>
1.417126MB/s 00000000000000000000>
1.496284MB/s 000000000000000000000>
1.442938MB/s 00000000000000000000>
1.743159MB/s 0000000000000000000000000>
1.716102MB/s 000000000000000000000000>
1.765166MB/s 0000000000000000000000000>
1.509863MB/s 000000000000000000000>
1.638345MB/s 00000000000000000000000>
1.953884MB/s 0000000000000000000000000000>
1.723031MB/s 000000000000000000000000>
1.917233MB/s 000000000000000000000000000>
1.474753MB/s 000000000000000000000>

(I have just sent ifspeed to this list)

So sustained througput on eth0 (Digital DE500) below 2MBytes/s currently.
used to be >5MBytes/s with this many users (they are either changing and
reading dirs, or reading or writing >500MBytes files, so speed,
bandwidth, throughput would be essential)

What I had w/2.2.18pre19 (+raid+ide):
~80MB more in cache and ~80MB swapped out (eg. currently unused notes
server and squid) There is enough of swap over 3 disks (like the
raid), so I did not bother disabling squid and notes, since - I thought -
they would only take up some swap unused.

Unfortunatelly not so.

Anybody has ideas of some /proc tuning maybe able to help my situation?

( currently I use:
# Enable more files and inodes to be used
echo '65536' > /proc/sys/fs/file-max
echo '262144' > /proc/sys/fs/inode-max
echo "10240 61000" > /proc/sys/net/ipv4/ip_local_port_range
echo "134217728" > /proc/sys/kernel/shmmax
echo "300" > /proc/sys/net/ipv4/tcp_keepalive_time
)
-- 
SaPE - Peter, Sasi - mailto:sape@sch.hu - http://sape.iq.rulez.org/


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: 2.2.19pre6aa1 degraded performance for me...

[Andrea Arcangeli]

On Mon, Jan 08, 2001 at 10:46:29PM +0100, Sasi Peter wrote:
> What I had w/2.2.18pre19 (+raid+ide):
> ~80MB more in cache and ~80MB swapped out (eg. currently unused notes
> server and squid) There is enough of swap over 3 disks (like the
> raid), so I did not bother disabling squid and notes, since - I thought -
> they would only take up some swap unused.

There are many variables. However I guess the slowdown is because your idle
apps didn't got swapped out in favour of cache as you noticed. An aggressive
aging algorithm would probably fix that but it then would hurt other cases
(after you don't need a frequenty accessed part of filesystem cache anymore it
would take ages before it gets collected potentially causing an unnecessary
swapout storms because the kernel doesn't know you don't need such cache
anymore).  Furthmore if notes and squid are rarely running but they provides
critical services if they would go totally into swap in favour of fs cache you
would get very bad latencies the first time somebody connects to the server. So
the fix I suggest you is to buy more ram or to shutdown squid and notes. Than
you may as well see a performance improvement compared to 2.2.18pre19
(+raid+ide).  Otherwise you can push the machine low on memory a bit until they
both goes totally into swap (check with `ps v`). Hope this helps.

Andrea
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: 2.2.19pre6aa1 degraded performance for me...

[Sasi Peter]

On Tue, 9 Jan 2001, Andrea Arcangeli wrote:

> On Mon, Jan 08, 2001 at 10:46:29PM +0100, Sasi Peter wrote:
> > What I had w/2.2.18pre19 (+raid+ide):
> > ~80MB more in cache and ~80MB swapped out (eg. currently unused notes
> > server and squid) There is enough of swap over 3 disks (like the
> > raid), so I did not bother disabling squid and notes, since - I thought -
> > they would only take up some swap unused.
>
> There are many variables. However I guess the slowdown is because your idle
> apps didn't got swapped out in favour of cache as you noticed. An aggressive
> aging algorithm would probably fix that but it then would hurt other cases
> (after you don't need a frequenty accessed part of filesystem cache anymore it
> would take ages before it gets collected potentially causing an unnecessary
> swapout storms because the kernel doesn't know you don't need such cache
> anymore).  Furthmore if notes and squid are rarely running but they provides
> critical services if they would go totally into swap in favour of fs cache you
> would get very bad latencies the first time somebody connects to the server. So
> the fix I suggest you is to buy more ram or to shutdown squid and notes. Than

Oh well I thought 384MB should be enought for everyone aiming at this
performance (almost TM ;). At least it would up till now :(

> you may as well see a performance improvement compared to 2.2.18pre19
> (+raid+ide).  Otherwise you can push the machine low on memory a bit until they
> both goes totally into swap (check with `ps v`). Hope this helps.

I'll try this, thanks. (so no echo '1 23 456' >/proc/sys/vm/...?)

-- 
SaPE - Peter, Sasi - mailto:sape@sch.hu - http://sape.iq.rulez.org/


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: 2.2.19pre6aa1 degraded performance for me...

[Sasi Peter]

On Tue, 9 Jan 2001, Sasi Peter wrote:
> On Tue, 9 Jan 2001, Andrea Arcangeli wrote:
> > On Mon, Jan 08, 2001 at 10:46:29PM +0100, Sasi Peter wrote:
> > > What I had w/2.2.18pre19 (+raid+ide):
> > > ~80MB more in cache and ~80MB swapped out (eg. currently unused notes
> > > server and squid) There is enough of swap over 3 disks (like the
> > > raid), so I did not bother disabling squid and notes, since - I thought -
> > > they would only take up some swap unused.
> > There are many variables. However I guess the slowdown is because your idle
> > apps didn't got swapped out in favour of cache as you noticed. An aggressive
> > aging algorithm would probably fix that but it then would hurt other cases
> > (after you don't need a frequenty accessed part of filesystem cache anymore it
> > would take ages before it gets collected potentially causing an unnecessary
> > swapout storms because the kernel doesn't know you don't need such cache
> > anymore).  Furthmore if notes and squid are rarely running but they provides
> > critical services if they would go totally into swap in favour of fs cache you
> > would get very bad latencies the first time somebody connects to the server. So
> > the fix I suggest you is to buy more ram or to shutdown squid and notes. Than
> Oh well I thought 384MB should be enought for everyone aiming at this
> performance (almost TM ;). At least it would up till now :(
> > you may as well see a performance improvement compared to 2.2.18pre19
> > (+raid+ide).  Otherwise you can push the machine low on memory a bit until they
> > both goes totally into swap (check with `ps v`). Hope this helps.
> I'll try this, thanks. (so no echo '1 23 456' >/proc/sys/vm/...?)

I thought it over again. I still have to say it is a nonsense for a kernel
not to  have _anything_ (zero pages) currently unused swapped out under
such an I/O load!

-- 
SaPE - Peter, Sasi - mailto:sape@sch.hu - http://sape.iq.rulez.org/


-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/

Re: 2.2.19pre6aa1 degraded performance for me...

[Andrea Arcangeli]

On Wed, Jan 10, 2001 at 01:07:55AM +0100, Sasi Peter wrote:
> I thought it over again. I still have to say it is a nonsense for a kernel
> not to  have _anything_ (zero pages) currently unused swapped out under
> such an I/O load!

Could you generate some furhter memory pressure to see what happens? Do you
confirm that you get the same performance as with your previous kernel if the
idle servers gets swapped out or if you shutdown them?

Andrea
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/