What is swapping ?

Question

I compare vmstat & top output:
e220:/ > vmstat -S 60
procs memory page disk faults cpu
r b w swap free si so pi po fr de sr s0 s1 s6 s2 in sy cs us sy id
0 0 35 2519760 272144 0 0 181 68 89 0 133 14 13 0 2 692 543 451 26 8 66
2 0 34 2398872 26216 0 0 120 196 202 0 101 17 17 0 3 1268 2853 909 88 12 0
2 0 34 2381056 24144 0 0 93 304 318 0 128 18 16 0 3 1202 2917 867 88 12 0
1 0 34 2379280 30064 0 0 149 153 160 0 62 18 15 0 3 909 2502 670 90 10 0
2 0 34 2379000 25144 0 0 145 296 307 0 132 17 19 0 3 1220 2895 893 88 12 0
top -M (sorted by memory) (only subset of full output of top)
PID USERNAME LWP PRI NICE SIZE RES STATE TIME CPU COMMAND
17479 ora9ias 14 58 0 100M 18M sleep 7:39 0.00% emagent
20675 ora9ias 5 58 0 92M 14M sleep 0:12 0.00% httpd
21131 ora9ias 5 58 0 89M 14M sleep 0:10 0.00% httpd
20762 ora9ias 5 52 0 87M 18M sleep 0:10 0.00% httpd
21151 ora9ias 5 58 0 85M 18M sleep 0:08 0.00% httpd
21461 ora9ias 5 58 0 83M 15M sleep 0:06 0.00% httpd
20589 ora9ias 5 58 0 76M 17M sleep 0:11 0.00% httpd
The w column in vmstat shows the number of swapped out processes.
But top shows the resident size and all size of process and there
no one process which is fully swaped out.
Therefore the question:
if 90% pages of process are in swap and 10% are in memory then could vmstat increments w?
if 10% pages of process are in swap and 90% are in memory then could vmstat increments w?
For example:
PID SIZE RES
17479 100M 95M
1031 100M 5M
what is the w ? 0 or 1 or 2 ?
If the process contains some LWP, then could any single LWP be swaped out or not?
Thanks!

Answer 1

Well, swapping out a process only removes all of its private pages. Any shared libraries that it's using (which could be a huge fraction of the total mappings) may not be flushed. Even if they were, they might be brought back in by another process shortly afterward.
So if you want to try looking at this further, I'd recommend looking at 'pmap -x <PID>' output.
0 0 35 2519760 272144 0 0 181 68 89 0 133 14 13 0 2 692 543 451 26 8 66
2 0 34 2398872 26216 0 0 120 196 202 0 101 17 17 0 3 1268 2853 909 88 12 0I looked at this a long time ago and couldn't make things line up. But one thing I wasn't aware of then is that the 'w' column does not represent processes, it represents threads. So when my 'w' column was 20 and I couldn't find 20 processes that looked completely paged out, maybe that was right and some of them had multiple threads.
The w column in vmstat shows the number of swapped out processes.
But top shows the resident size and all size of process and there
no one process which is fully swaped out.Right, because shared pages are counted in RES. (That's why you can't add all the RES figures to get the total amount of memory consumption. You'd be counting some pages multiple times).
Therefore the question:
if 90% pages of process are in swap and 10% are in memory then could vmstat increments w?
if 10% pages of process are in swap and 90% are in memory then could vmstat increments w?
For example:
PID SIZE RES
17479 100M 95M
1031 100M 5M
what is the w ? 0 or 1 or 2 ?Indeterminate for two reasons:
1) A process could have any of its pages sent to backing store through normal paging processes. Even if they all leave this way, the process has not been "swapped".
2) When a process is swapped, shared pages will remain.
If the process contains some LWP, then could any single LWP be swaped out or not?My understanding is no. "Swapping" is a memory managment thing, and all threads share the same memory. So swapping a process will send all memory to disk, affecting all threads. But I haven't confirmed that (still, no one has contradicted me about it either).
Darren