Re: [2.4.17/18pre] VM and swap - it's really unusable

[Ken Brownfield <brownfld@irridia.com>]

On Mon, Feb 25, 2002 at 03:51:43AM -0800, Aviv Shavit wrote:
| Ken Brownfield's A) and B) hit me
| regularly(http://www.uwsg.indiana.edu/hypermail/linux/kernel/0201.0/0740.html)
| (Thanks Ken for starting this enlightening thread and
| to all those that contributed)

No problem.  You might be coming in late, but nothing's changed. :-/
Although the APIC thing might be MPS1.4 related, I'm finding.  I don't
remember if that thread had my workaround to the APIC issue, so I'll
attach that if you're feeling adventurous.  Works flawlessly for me in
production, for the last few months.

| Running software I developed:
| - running on 2.4.17
| - accessing a large number of files
| - 2GB memory
| - large multiple partitions - ext2
| 
| I saw references to patches by Martin and
| M.H.vanLeeuwen on this thread. Where can I get those
| (hopefully with a bit of info) ?

I've attached his patch; hopefully he doesn't mind.  His patch is
actually similar to code that's in rmap currently.

It does seem to temper the effects of [id]cache bloat for some of my
more common load patterns.  But even with that patch, under VM load the
mainline kernel collapses.  For instance, I have a large parallel task
that takes 5 minutes under -rmap or -aa, but I get bored and kill it
after 4 *hours* on mainline.  It dips into swap, but only by about 50MB.

Also, -aa doesn't seem to make as large an impact on shrinking the
[id]caches.  I imagine that without returning values,
shrink_[id]cache_memory() behavior is difficult to tune appropriately.

In any case, rmap-12f has been running fine for me.  I'm going to create
three kernels to distribute in production -- 2.4.18 more or less
vanilla, 2.4.18 with random debugging (thanks Andreas) to try to resolve
the /dev/random death issue, and 2.4.18+O(1)K3+rmap12f.

rmap is a tremendous improvement over mainline -- really the only data I
still need is stability, both of O(1) and rmap.  Porting Andrea's 10_vm
was a pain the last time I did it, and it didn't have all of the
positives of rmap.

Clearly I'm not in the majority, though.  While I get bitten on a weekly
basis by the 2.4 VM, very few other people mention it.  Maybe they just
assume that it's normal behavior?  Scary.

Anyway, short of the long is that I would suggest rmap if you're having
problems.  Rmap actually obviates three other patches I typically apply.
But be aware that rmap is still a bit of a work in progress, at least in
terms of tuning.

Cheers,
-- 
Ken.
brownfld@irridia.com


| I also saw references to 'rmap' on the 2.4.18
| changelog. Is that related ?
| 
| pls. cc' me on your posts back.
| 
| Thanks
| Aviv
| 
| "And all this science I don't understand,
|       It's just my job five days a week"
| Rocket Man, E.J.


--- linux.virgin/mm/vmscan.c	Mon Dec 31 12:46:25 2001
+++ linux/mm/vmscan.c	Fri Jan 11 18:03:05 2002
@@ -394,9 +394,9 @@
 		if (PageDirty(page) && is_page_cache_freeable(page) && page->mapping) {
 			/*
 			 * It is not critical here to write it only if
-			 * the page is unmapped beause any direct writer
+			 * the page is unmapped because any direct writer
 			 * like O_DIRECT would set the PG_dirty bitflag
-			 * on the phisical page after having successfully
+			 * on the physical page after having successfully
 			 * pinned it and after the I/O to the page is finished,
 			 * so the direct writes to the page cannot get lost.
 			 */
@@ -480,11 +480,14 @@
 
 			/*
 			 * Alert! We've found too many mapped pages on the
-			 * inactive list, so we start swapping out now!
+			 * inactive list.
+			 * Move referenced pages to the active list.
 			 */
-			spin_unlock(&pagemap_lru_lock);
-			swap_out(priority, gfp_mask, classzone);
-			return nr_pages;
+			if (PageReferenced(page) && !PageLocked(page)) {
+				del_page_from_inactive_list(page);
+				add_page_to_active_list(page);
+			}
+			continue;
 		}
 
 		/*
@@ -521,6 +524,9 @@
 	}
 	spin_unlock(&pagemap_lru_lock);
 
+	if (max_mapped <= 0 && (nr_pages > 0 || priority < DEF_PRIORITY))
+		swap_out(priority, gfp_mask, classzone);
+
 	return nr_pages;
 }