Skip to contentSkip to conversation

2.4.13pre5aa1

DISCUSSION Andrea Arcangeli 3 participants 5 messages Oct 19, 2001 – Oct 22, 2001

lore
Share & tools
lore
#0 lore raw

The vm part in particular is right now getting stressed on a 16G box kindly provided by osdlab.org and it didn't exibith any problem yet. This is a trace of the workload that is running on the machine overnight.

   procs                      memory    swap          io     system         cpu
 r  b  w   swpd   free   buff  cache  si  so    bi    bo   in    cs  us  sy  id
 0  3  2 7055840   5592    196   4208 681 903   687   904   56    87   0   5  95
 0  3  0 7055332   4956    184   4196 5892 4820  5892  4832  418   547   0   3  97
 0  3  0 7056004   5816    176   4184 6172 6400  6172  6400  418   579   0   4  96
 0  3  0 7055912   5688    192   4184 4720 4096  4736  4112  355   456   0   2  98
 0  3  1 7055852   5300    180   4220 5624 5068  5720  5072  408   526   0   2  98
 0  3  0 7055992   5228    176   4220 6384 5744  6384  5744  427   586   0   1  99
 0  3  0 7055900   5232    176   4220 6016 5676  6016  5676  417   545   0   2  98
 0  3  0 7056396   5844    180   4220 5644 5656  5656  5660  402   560   0   1  99
 0  3  1 7056476   6012    176   4216 6104 6144  6104  6144  411   582   0   1  99
 0  3  0 7056084   5592    176   4220 5540 4452  5540  4452  386   525   0   1  98
 0  3  0 7055948   5400    176   4220 5184 4724  5184  4724  355   519   0   2  98
 0  3  0 7056676   6136    176   4232 7360 7592  7360  7592  519   720   0   1  98
 0  4  0 7056264   5572    176   4240 5888 5112  5908  5112  411   535   0   1  99
 0  3  1 7055948   5444    180   4240 5632 4912  5656  4932  402   605   0   1  99
 0  3  0 7055780   5088    176   4240 4932 4276  4932  4276  350   432   0   1  99
 0  3  0 7055612   5128    176   4240 4564 4252  4564  4252  340   434   0   0 100
             total       used       free     shared    buffers     cached
Mem:      16493180   16488716       4464          0        184       4260
-/+ buffers/cache:   16484272       8908

Swap: 36941584 7054904 29886680

It seems still very responsive despite of the load (and also despite being at the other side of the Atlantic :).

URL:

	ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.13pre5aa1.bz2
	ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.13pre5aa1/

If it swaps too much please try:

echo 6 > /proc/sys/vm/vm_scan_ratio
echo 2 > /proc/sys/vm/vm_mapped_ratio
echo 4 > /proc/sys/vm/vm_balance_ratio

Thanks!

--
Only in 2.4.13pre5aa1: 00_alpha-rest-pci-1
Only in 2.4.13pre5aa1: 00_alpha-tsunami-1

iommu alpha fixes from Jay Estabrook, Ivan Kokshaysky and Richard Henderson.

Only in 2.4.13pre3aa1: 00_files_struct_rcu-2.4.10-04-1 Only in 2.4.13pre5aa1: 00_files_struct_rcu-2.4.10-04-2

Latest uptdate from Maneesh Soni including the memalloc faliure bugfix Chip Salzenberg.

Only in 2.4.13pre3aa1: 00_highmem-deadlock-1
Only in 2.4.13pre5aa1: 00_highmem-deadlock-2

Rediffed so that it's self contained (previously it wasn't very readable).

Only in 2.4.13pre3aa1: 00_lowlatency-fixes-1
Only in 2.4.13pre5aa1: 00_lowlatency-fixes-2

	Added a reschedule point, mainly for madvise but it's in the ->nopage
	way too.

Only in 2.4.13pre3aa1: 00_seg-reload-1

Dropped, the common case is user<->kernel, and it have to be very fast too since it's more important. (idling routers would better not use irq at all but to dedicate a cpu to the polling work)

Only in 2.4.13pre3aa1: 00_vm-3
Only in 2.4.13pre3aa1: 00_vm-3.2
Only in 2.4.13pre5aa1: 10_vm-4

Further vm work. Not sure if this is better than the previous one, but now the few magic numbers are sysctl configurable. Included many fixes from Linus (that are also in pre5 of course) and one fix from Manfred to avoid interrupt to eat the pfmemalloc reserved pool.

Only in 2.4.13pre3aa1: 10_highmem-debug-5
Only in 2.4.13pre5aa1: 20_highmem-debug-6

Fixed zone alignment to avoid crashes when highmem emulation is enabled.

Only in 2.4.13pre5aa1: 10_lvm-deadlock-fix-1

	Dropped sync_dev from blkdev ->close to avoid a deadlock on lvm close(2),
	fix from Alexander Viro.

Only in 2.4.13pre3aa1: 60_tux-2.4.10-ac12-H1.bz2
Only in 2.4.13pre5aa1: 60_tux-2.4.10-ac12-H8.bz2

Latest update from Ingo Molnar at www.redhat.com/~mingo/ .
--

Andrea

#1 lore raw
Quoted 2 lines · Andrea Arcangeli

Only in 2.4.13pre3aa1: 00_files_struct_rcu-2.4.10-04-1 Only in 2.4.13pre5aa1: 00_files_struct_rcu-2.4.10-04-2

I want to point out to preempt-kernel users that RCU is not preempt-safe. The implicit locking assumed from per-CPU data structures is defeated by preemptibility.

(Actually, FWIW, I think I can think of ways to make RCU preemptible but it would involve changing the write-side quiescent code for the case where the pointers were carried over the task switches. Probably not worth it.)

This is not to say RCU is worthless with a preemptible kernel, but that we need to make it safe (and then make sure it is still a performance advantage, but I don't think this would add much overhead). Note this is clean, simply wrapping the read code in non-preemption statements.

I will hack up a patch when I get the time, but I would like to prevent myself from maintaining the patch against a third tree ... where, oh where, is 2.5? :)

Robert Love

#2 lore raw
In reply to Andrea Arcangeli
Quoted 3 lines · Andrea Arcangeli

The vm part in particular is right now getting stressed on a 16G box kindly provided by osdlab.org and it didn't exibith any problem yet. This is a trace of the workload that is running on the machine overnight.

Hello Andrea,

I did some performance comparison on my machine (Athlon 1200, 256MB DDR, IDE hdd). Basically all I did was time the complete build prozess of a kernel:

#!/bin/bash

REV=uname -r-date +%s
LF="/usr/src/logfile-$REV"
PAR=100

tar xvIf /home/public/Linux/kernel/v2.4/linux-2.4.12.tar.bz2 >>$LF 2>&1 cd linux
bzip2 -cd /home/public/Linux/kernel/v2.4/patch-2.4.13-pre5.bz2 | patch -p1

Quoted 1 line
Quoted 1 line

$LF 2>&1

bzip2 -cd /home/public/Linux/kernel/v2.4/2.4.13pre5aa1.bz2 | patch -p1 >>$LF 2>&1
patch -p0 < ../Makefile.patch >>$LF 2>&1
cp ../config-2.4.13-pre5aa1 .config
(make oldconfig dep clean && make -j$PAR bzImage modules && cat /proc/loadavg && cat /proc/meminfo) >

Quoted 1 line

$LF 2>&1

I ran the above script five times in a row after a fresh reboot into single user mode so that no other processes could interfere. Here are the results:

                    j25       j50       j75       j100
2.4.12-ac3:       4:52.16   5:21.15   6:22.85   10:26.70
2.4.12-ac3:       4:52.31   5:22.52   8:40.97   15:55.37
2.4.12-ac3:       4:51.51   5:18.04   6:08.80   7:09.85
2.4.12-ac3:       4:52.47   5:10.81   5:51.02   8:53.40
2.4.12-ac3:       4:53.24   5:06.96   6:36.49   7:53.92
2.4.13-pre3aa1:   4:52.08   5:57.43   9:43.07      *
2.4.13-pre3aa1:   4:54.22   5:04.17   10:30.56     *
2.4.13-pre3aa1:   4:53.95   5:21.08   11:07.44     *
2.4.13-pre3aa1:   4:55.26   5:30.01   9:53.16      *
2.4.13-pre3aa1:   4:54.39   6:00.32   10:15.06     *
2.4.13-pre5aa1:   4:54.61   5:10.38   5:19.68   5:40.37
2.4.13-pre5aa1:   4:56.39   5:12.39   5:31.16   5:59.54
2.4.13-pre5aa1:   4:55.12   5:11.00   5:13.37   5:43.99
2.4.13-pre5aa1:   4:56.24   5:10.85   5:16.17   5:50.78
2.4.13-pre5aa1:   4:57.05   5:10.97   5:26.41   6:06.42

* Kernel build did not complete because OOM killer was activated.

If further infos are interesting just send me an email.

Kind regards,

Jochen

Signature · 3 lines
Well, yeah ... I suppose there's no point in getting greedy, is there?

    << Calvin & Hobbes >>
#3 lore raw
Quoted 10 lines · Rik van Riel
Quoted 3 lines · On 21 Oct 2001 [email protected] wrote:

2.4.12-ac3: 4:52.16 5:21.15 6:22.85 10:26.70

If further infos are interesting just send me an email.

It would be cool if you could test this with 2.4.12-ac3 and my -vmpatch and -freeswap patches against this kernel ;)

Patches on http://www.surriel.com/patches/

As promised here are the results for 2.4.12-ac3 + vmpatch and freeswap.

2.4.12-ac3: 8:20.46
2.4.12-ac3: 6:36.67
2.4.12-ac3: 6:37.61
2.4.12-ac3: 7:36.24
2.4.12-ac3: 7:21.24

These times are for -j100.

Kind regards,

Jogi

Signature · 3 lines
Well, yeah ... I suppose there's no point in getting greedy, is there?

    << Calvin & Hobbes >>
#4 lore raw
In reply to Andrea Arcangeli
Quoted 1 line · On Sun, Oct 21, 2001 at 09:17:26PM +0200, [email protected] wrote:

2.4.13-pre3aa1: 4:54.39 6:00.32 10:15.06 *

					^^^^^^^^
Quoted 1 line

2.4.13-pre5aa1: 4:54.61 5:10.38 5:19.68 5:40.37

					^^^^^^^

this is interesting. I'm also wondering what you'd get if you used:

echo 8 > /proc/sys/vm/vm_scan_ratio
echo 1 > /proc/sys/vm/vm_mapped_ratio
echo 3 > /proc/sys/vm/vm_balance_ratio

(or also the other combination that I suggested in the other emails)

Anyways you can probably skip the above test and wait for a further update that changes more than just the default sysctl values (also notably it introduces the PG_launder logic originated from a discussion with Marcelo and Linus, resemling somehow part of the PG_wait_for_IO write throttling logic that I had in 2.4.12aa1 and 2.4.13pre3aa1, but I doubt pre3aa1 was slower because of that, and in case next -aa will slowdown again I'll later ask you to try with a one liner patch that will disable the write throttling for writepage again [like pre5aa1 did] just to make sure it's not the one that hurts :)

thanks to you too for the feedback!

Andrea