lkml.org 
[lkml]   [2005]   [Oct]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH 0/4] Swap migration V3: Overview
On 10/24/05, Marcelo Tosatti <marcelo.tosatti@cyclades.com> wrote:
> On Sun, Oct 23, 2005 at 09:50:18PM +0900, Magnus Damm wrote:
> > Maybe SLAB defragmentation code is suitable for page migration too?
>
> Free dentries are possible to migrate, but not referenced ones.
>
> How are you going to inform users that the address of a dentry has
> changed?

Um, not sure, but the idea of defragmenting SLAB entries might be
similar to moving them, ie migration. But how to solve the per-SLAB
referencing is another story... =)

> > > > But I'm probably underestimating the cost of page migration...
> > >
> > > The zone balancing issue you describe might be an issue once zone
> > > said pages can be migrated :)
> >
> > My main concern is that we use one LRU per zone, and I suspect that
> > this design might be suboptimal if the sizes of the zones differs
> > much. But I have no numbers.
>
> Migrating user pages from lowmem to highmem under situations with
> intense low memory pressure (due to certain important allocations
> which are restricted to lowmem) might be very useful.

I patched the kernel on my desktop machine to provide some numbers.
The zoneinfo file and a small patch is attached.

$ uname -r
2.6.14-rc5-git3

$ uptime
20:27:47 up 1 day, 6:27, 18 users, load average: 0.01, 0.13, 0.15

$ cat /proc/zoneinfo | grep present
present 4096
present 225280
present 30342

$ cat /proc/zoneinfo | grep tscanned
tscanned 151352
tscanned 3480599
tscanned 541466

"tscanned" counts how many pages that has been scanned in each zone
since power on. Executive summary assuming that only LRU pages exist
in the zone:

DMA: each page has been scanned ~37 times
Normal: each page has been scanned ~15 times
HighMem: each page has been scanned ~18 times

So if your user space page happens to be allocated from the DMA zone,
it looks like it is more probable that it will be paged out sooner
than if it was allocated from another zone. And this is on a half year
old P4 system.

> > There are probably not that many drivers using the DMA zone on a
> > modern PC, so instead of bringing performance penalty on the entire
> > system I think it would be nicer to punish the evil hardware instead.
>
> Agreed - the 16MB DMA zone is silly. Would love to see it go away...

But is the DMA zone itself evil, or just that we have one LRU per zone...?

/ magnus
[unhandled content-type:application/octet-stream]--- from-0002/include/linux/mmzone.h
+++ to-work/include/linux/mmzone.h 2005-10-24 10:43:13.000000000 +0900
@@ -151,6 +151,7 @@ struct zone {
unsigned long nr_active;
unsigned long nr_inactive;
unsigned long pages_scanned; /* since last reclaim */
+ unsigned long pages_scanned_total;
int all_unreclaimable; /* All pages pinned */

/*
--- from-0002/mm/page_alloc.c
+++ to-work/mm/page_alloc.c 2005-10-24 10:51:05.000000000 +0900
@@ -2101,6 +2101,7 @@ static int zoneinfo_show(struct seq_file
"\n active %lu"
"\n inactive %lu"
"\n scanned %lu (a: %lu i: %lu)"
+ "\n tscanned %lu"
"\n spanned %lu"
"\n present %lu",
zone->free_pages,
@@ -2111,6 +2112,7 @@ static int zoneinfo_show(struct seq_file
zone->nr_inactive,
zone->pages_scanned,
zone->nr_scan_active, zone->nr_scan_inactive,
+ zone->pages_scanned_total,
zone->spanned_pages,
zone->present_pages);
seq_printf(m,
--- from-0002/mm/vmscan.c
+++ to-work/mm/vmscan.c 2005-10-24 10:44:09.000000000 +0900
@@ -633,6 +633,7 @@ static void shrink_cache(struct zone *zo
&page_list, &nr_scan);
zone->nr_inactive -= nr_taken;
zone->pages_scanned += nr_scan;
+ zone->pages_scanned_total += nr_scan;
spin_unlock_irq(&zone->lru_lock);

if (nr_taken == 0)
@@ -713,6 +714,7 @@ refill_inactive_zone(struct zone *zone,
pgmoved = isolate_lru_pages(nr_pages, &zone->active_list,
&l_hold, &pgscanned);
zone->pages_scanned += pgscanned;
+ zone->pages_scanned_total += pgscanned;
zone->nr_active -= pgmoved;
spin_unlock_irq(&zone->lru_lock);
\
 
 \ /
  Last update: 2005-10-25 13:41    [W:0.496 / U:0.188 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site