VM regression in 2.6.7-bk20

[Shantanu Goel]

Hi,

I believe there is VM regression due to the recent
changes to the page scanner.  It stems from the fact
that shrink_zone() limits # pages reclaimed in one go
to SWAP_CLUSTER_MAX instead of basing it on the memory
shortfall which causes increased pressure on the lower
zones.  I first noticed this while running my normal
workstation workload.  The machine seemed to swap more
than I expected because the # mapped pages was only
about 30%.  I noticed the DMA zone was being scanned
very frequently while running a script that monitored
/proc/vmstat. The "swappiness" value was the default
60.  Unfortunately, I failed to save the output so if
the statistics below are insufficient evidence, I can
reboot the stock kernel and get the output.

Below are the VM statistics for stock 2.6.7-bk20 and
modified 2.6.7-bk20 which sets sc.nr_to_reclaim to
zone->pages_high - zone->free_pages.  The statistics
were collected by a perl script that read /proc/vmstat
before and after running the benchmark and printing
the difference.  "swappiness" was again the default
value 60 in both cases.  I used kernbench with options
"-H -M" using the source for 2.6.0 so only the optimal
load (-j16) was run.  The machine has 2x2.0Ghz Xeons
with HT enabled and I manually booted it with mem=256M
though it has 1GB of memory.

                         modified     stock
pgscan_kswapd_normal     750420       655116
pgscan_direct_normal     561924       497640
pgscan_kswapd_dma        121903       390318
pgscan_direct_dma         78258       153638
pgsteal_normal           631517       575534
pgsteal_dma               67128       125950

Notice that the stock kernel scanned the DMA zone
about 3 times as often than the modified one, and
allocated twice as much memory from the DMA zone.

The modification I made is attached below.  This is
not meant for inclusion as the sysctls (limit_reclaim,
slow_scan) are for testing only.  Please ignore the
slow_scan stuff.  It was disabled during the tests
above and is something I was playing with on the side.
 The scanner used the stock algorithm in both cases
for deciding how much of the active/inactive lists to
scan.
I used limit_reclaim to enable/disable the stock
reclaim behaviour.

Thanks,
Shantanu



		
__________________________________
Do you Yahoo!?
New and Improved Yahoo! Mail - Send 10MB messages!
http://promotions.yahoo.com/new_mail

Attachment: vmscan.patch
Description: vmscan.patch