Skip to contentSkip to conversation

2.4.17pre1aa1

DISCUSSION Andrea Arcangeli 3 participants 3 messages Nov 29, 2001 – Nov 30, 2001

lore
Share & tools
lore
#0 lore raw

URL:

	ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.17pre1aa1.bz2
	ftp://ftp.us.kernel.org/pub/linux/kernel/people/andrea/kernels/v2.4/2.4.17pre1aa1/

Only in 2.4.17pre1aa1: 00_blkdev-ulimit-1

Avoid SIGXFSZ with old libc or with ulimit set on a blkdev.

Only in 2.4.15aa1: 00_iput-unmount-corruption-fix-1 Only in 2.4.15aa1: 00_read_super-stale-inode-1

Use fix in mainline (equivalent in practice).

Only in 2.4.15aa1: 00_silent-stack-overflow-10 Only in 2.4.17pre1aa1: 00_silent-stack-overflow-11

Rediffed.

Only in 2.4.17pre1aa1: 00_time_vs_gettimeofday-1

Make time() monotone against gettimeofday (from Andi Kleen).

Only in 2.4.15aa1: 10_block-highmem-all-18b-1 Only in 2.4.17pre1aa1: 10_block-highmem-all-18b-2

Alternate fix for the bounce_pfn boundary check (avoid bouncing one more page than necessary sometime).

Only in 2.4.15aa1: 10_vm-17
Only in 2.4.17pre1aa1: 10_vm-18

	Minor vm tweaks in function of the feedback received.
	Included Andrews' dirty += BUF_LOCKED.

Only in 2.4.15aa1: 50_uml-patch-2.4.14-5.bz2 Only in 2.4.17pre1aa1: 50_uml-patch-2.4.15-3.bz2

Latest update from Jeff.

Only in 2.4.17pre1aa1: 60_tux-flush_icache_range-1

Needed in -aa.

Only in 2.4.15aa1: 60_tux-2.4.15-pre9-B1.bz2 Only in 2.4.17pre1aa1: 60_tux-2.4.16-final-A3.bz2

Latest update from Ingo.

Andrea

#1 lore raw
Quoted 7 lines · Andrea Arcangeli

Only in 2.4.15aa1: 10_vm-17
Only in 2.4.17pre1aa1: 10_vm-18

Minor vm tweaks in function of the feedback received. Included Andrews' dirty += BUF_LOCKED.

OK. One think I notice is that you've also decreased nfract,nfract_sync from (40%,60%) to (20%,40%). So taken together, these changes mean that we'll start writeout much earlier, and will block writers much earlier. What's the thinking here?

I received some interesting results from Mike Galbraith today. Quoted without permission...

Quoted 39 lines

2.5.1-pre1

This was initial test of effect on throughput at generic page mover load.. parallel make of kernel.

real    7m54.873s
user    6m41.070s
sys     0m30.170s
user  :       0:06:47.35  72.6%  page in :   661891
nice  :       0:00:00.00   0.0%  page out:   708836
system:       0:00:47.42   8.5%  swap in :   140234
idle  :       0:01:46.26  18.9%  swap out:   172775

2.5.1-pre1+vm-fixes

real    7m48.438s
user    6m41.070s
sys     0m29.570s
user  :       0:06:47.89  74.9%  page in :   666952
nice  :       0:00:00.00   0.0%  page out:   621296
system:       0:00:47.70   8.8%  swap in :   142391
idle  :       0:01:28.94  16.3%  swap out:   150721 * (free)

(very interesting imho.. particularly idle time)

2.5.1-pre1+vm-fixes+elevator

real    8m13.386s
user    6m38.330s
sys     0m31.680s
user  :       0:06:45.24  70.3%  page in :   596724
nice  :       0:00:00.00   0.0%  page out:   574456
system:       0:00:47.79   8.3%  swap in :   123507
idle  :       0:02:03.64  21.4%  swap out:   138675

(free for this load)

2.5.1-pre1+vm-fixes+elevator+mini-ll

Quoted 8 lines
real    8m12.437s
user    6m38.860s
sys     0m31.680s
user  :       0:06:45.90  71.0%  page in :   604385
nice  :       0:00:00.00   0.0%  page out:   572588
system:       0:00:47.50   8.3%  swap in :   126731
idle  :       0:01:58.05  20.7%  swap out:   138055

So we see that the dirty += BUF_LOCKED thing appears to increase the parallel-make-on-64meg-machine workload.

Unfortunately the vm-fixes patch also has a few tweaks to decrease swapout and eviction, and we can see from Mike's numbers that the page in/out rate has dropped quite a bit. So we don't know which change caused the (rather modest) throughput improvement.

We also see that the elevator read latency improvements have caused a 5%-6% drop in throughput, which is to be expected. Tuning that back with `elvtune -b' will presumably get the aggregate throughput back, at the expense of interactivity.

Generally, your VM patch is getting really, really large, Andrea. This is a bit awkward, because we're twiddling so many knobs at the same time, and there's not a lot of description about what all the bits do. Is it possible to break it up
into smaller units? What are your plans for sending this
patch to Marcelo?

Thanks.

-

#2 lore raw
Quoted 9 lines · Andrew Morton
Quoted 7 lines · Andrea Arcangeli

Only in 2.4.15aa1: 10_vm-17
Only in 2.4.17pre1aa1: 10_vm-18

Minor vm tweaks in function of the feedback received. Included Andrews' dirty += BUF_LOCKED.

Quoted 6 lines

OK. One think I notice is that you've also decreased nfract,nfract_sync from (40%,60%) to (20%,40%). So taken together, these changes mean that we'll start writeout much earlier, and will block writers much earlier. What's the thinking here?

I received some interesting results from Mike Galbraith today.

2.4.17-pre1aa1

real    7m39.066s
user    6m38.400s
sys     0m29.140s
user  :       0:06:44.82  76.1%  page in :   536247
nice  :       0:00:00.00   0.0%  page out:   466800
system:       0:00:45.45   8.5%  swap in :   106783
idle  :       0:01:21.89  15.4%  swap out:   111683

__alloc_pages: 0-order allocation failed (gfp=0xf0/0) (only the one)

IO contrast...

2.5.1-pre1

real    7m54.873s
user    6m41.070s
sys     0m30.170s
user  :       0:06:47.35  72.6%  page in :   661891
nice  :       0:00:00.00   0.0%  page out:   708836
system:       0:00:47.42   8.5%  swap in :   140234
idle  :       0:01:46.26  18.9%  swap out:   172775

-Mike