kswapd is having a party

[Duraid Madina]

My question for today: does anyone know what kswapd actually does?

This morning, my workstation (linux ia64 2.6.0-test3) was fairly lightly 
loaded:

08:28:50 up 2 days,  1:49,  4 users,  load average: 1.77, 1.79, 1.20
Mem:   2062032k total,  1916368k used,   145664k free,   119952k buffers
Swap:  2016112k total,     1984k used,  2014128k free,   697920k cached

As you can see, there was gobs of RAM I wasn't using.

Anyway, as I was enjoying my tasty breakfast cereal, I noticed kswapd 
decided to wake up and say hello:

29109 duraid    25   0  400m 394m 6272 R 95.0 19.6   3:07.09 lsh
29108 duraid    19   0  404m 399m 6272 R 48.3 19.8   3:15.89 lsh
   13 root      15   0     0    0    0 R 17.2  0.0   0:26.47 kswapd0

29148 duraid    25   0  400m 394m 6272 R 93.9 19.6   3:13.23 lsh
29147 duraid    20   0  404m 399m 6272 S 21.1 19.8   3:12.16 lsh
   13 root      15   0     0    0    0 R 18.1  0.0   0:38.02 kswapd0

29148 duraid    25   0  400m 394m 6272 R 75.9 19.6   5:32.66 lsh
29147 duraid    18   0  404m 399m 6272 S 62.0 19.8   4:08.49 lsh
   13 root      16   0     0    0    0 R 16.4  0.0   0:50.08 kswapd0

29147 duraid    25   0  404m 397m 6272 R 93.9 19.7   6:23.87 lsh
29148 duraid    25   0  400m 392m 6272 R 75.4 19.5   9:00.72 lsh
   13 root      15   0     0    0    0 R 21.0  0.0   1:03.84 kswapd0

	It seems I can make kswapd use fairly nontrivial amounts of CPU time by 
running a perfectly innocent userland program. lsh is a raytracer I'm 
starting to work on. It's a standard C program, which uses MPI (I'm 
using LAM 7.0). The parallelism is trivial - each process renders a 
section of the image (divided vertically), line-at-a-time. The imbalance 
in CPU time between the processes is due to some sections of the scene 
being less interesting than others and finishing more quickly.

	Does anyone have _any_ idea what kswapd might actually be doing? I 
checked: not a single page was swapped in our out througout the duration 
of this test. Is there a chance that spinning on some lock (I have no 
idea how LAM does its synchronization), or perhaps even just idling, 
might be counted as kswapd0?

	Curious,

	Duraid

Attached: a snap of gkrellm. the light blue is kswapd, the dark blue is 
my program, the black is idle time.