Encouraging results. The very non-scientific benchmark is, boot with 8Mb,
allocate 100Mb and sequentially access every page in a loop.
On the first 100Mb scan, we are only swapping data out to disk. Both 2.0
and 2.1 manage to boot stuff out to swap at ~5Mb/s. 2.1 _is_ a little
faster but not much. My disk is ~6.5Mb/s raw read speed.
However on the 2nd 100Mb scan we are both reading in from swap and paging
to disk simultaneously.
2.0 manages ~0.4Mb/s (so 0.8Mb/s raw disk activity)
2.1.131ac8 can top 1.5Mb/s (>3Mb/s raw disk activity)
This is pretty good and makes Linux much more suitable for scientific
applications which make linear sweeps through huge arrays. I may
experiment with smaller swap space to see if this reduces seek time and
improves performance.
I'll expermiment with Stephens latest patch later too.
Just one strange anomaly. After trying the same test with 64Mb and killing
the program, about 8Mb of swap was tagged as still used. However there
certainly wasn't 8Mb of stuff in total.
Curious, I did swapoff -a. The 8Mb of swap disappeared (obviously) but my
"cached" figure also dropped 8Mb. Are we leaking cached stuff into swap??
Chris
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.rutgers.edu
Please read the FAQ at http://www.tux.org/lkml/