Re: 2.4.12aa1 [was Re: 2.4.11aa1 [was Re: 2.4.11pre6aa1]]

[Andrea Arcangeli <andrea@suse.de>]

On Thu, Oct 11, 2001 at 09:59:17PM +0200, Lorenzo Allegrucci wrote:
> Linux-2.4.11:
> 
> lenstra:~/src/qsort> time ./qsbench -n 90000000 -p 1 -s 140175100
> seed = 140175100
> 71.020u 1.650s 2:20.74 51.6%    0+0k 0+0io 10652pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 90000000 -p 1 -s 140175100
> seed = 140175100
> 71.070u 1.650s 2:21.51 51.3%    0+0k 0+0io 10499pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 90000000 -p 1 -s 140175100
> seed = 140175100
> 70.790u 1.670s 2:21.01 51.3%    0+0k 0+0io 10641pf+0w
> 
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 63.500u 1.800s 1:34.06 69.4%    0+0k 0+0io 8836pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 63.020u 1.470s 1:20.22 80.3%    0+0k 0+0io 6394pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 63.130u 1.520s 1:12.21 89.5%    0+0k 0+0io 5676pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 62.820u 1.560s 1:12.61 88.6%    0+0k 0+0io 5433pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 63.070u 1.560s 1:14.83 86.3%    0+0k 0+0io 5811pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 63.160u 1.650s 1:17.84 83.2%    0+0k 0+0io 6036pf+0w
> 
> lenstra:~/src/qsort> time ./qsbench -n 45000000 -p 2 -s 140175100
> 71.290u 2.010s 1:50.11 66.5%    0+0k 0+0io 10462pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 45000000 -p 2 -s 140175100
> 71.490u 2.220s 1:49.62 67.2%    0+0k 0+0io 11413pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 45000000 -p 2 -s 140175100
> 71.280u 2.360s 1:54.79 64.1%    0+0k 0+0io 11110pf+0w
> 
> 
> 
> Linux-2.4.12-aa1:
> 
> lenstra:~/src/qsort> time ./qsbench -n 90000000 -p 1 -s 140175100
> seed = 140175100
> 71.420u 2.110s 2:49.01 43.5%    0+0k 0+0io 16110pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 90000000 -p 1 -s 140175100
> seed = 140175100
> 70.960u 1.850s 2:45.05 44.1%    0+0k 0+0io 15463pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 90000000 -p 1 -s 140175100
> seed = 140175100
> 70.760u 1.980s 2:45.61 43.9%    0+0k 0+0io 15595pf+0w
> 
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 64.020u 1.940s 1:38.76 66.7%    0+0k 0+0io 10206pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 64.190u 1.410s 1:16.98 85.2%    0+0k 0+0io 6796pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 63.530u 1.520s 1:13.51 88.4%    0+0k 0+0io 5274pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 63.980u 1.370s 1:16.53 85.3%    0+0k 0+0io 6456pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 63.640u 1.680s 1:16.38 85.5%    0+0k 0+0io 6189pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 9000000 -p 10 -s 140175100
> 63.720u 1.500s 1:15.33 86.5%    0+0k 0+0io 5777pf+0w
> 
> lenstra:~/src/qsort> time ./qsbench -n 45000000 -p 2 -s 140175100
> 72.810u 2.010s 1:58.16 63.3%    0+0k 0+0io 14220pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 45000000 -p 2 -s 140175100
> 71.700u 2.290s 1:58.68 62.3%    0+0k 0+0io 13803pf+0w
> lenstra:~/src/qsort> time ./qsbench -n 45000000 -p 2 -s 140175100
> 72.440u 2.220s 1:56.13 64.2%    0+0k 0+0io 12911pf+0w

ok, it probably swapouts a little slower during the most heavy run,
that's pretty much intentional with the page-layer write throttling. The
little slowdown in swapout performance in the most heavy swapout
testcase should provide a much smoother responsiveness to the system
during the load that is more important than pure swap bandwith (if the
system feels unresponsive during the swap storm is much worse than the
pure speed of the allocations). In my interactive tests it was taking
more than 2 minutes to startup netscape during very heavy swapout
storms.  With this new code it took more than 1 minute showing a
sensible improvement in the understanding of the working set and the
swap beandwith decrease doesn't seem significant given that the cpu load
decreased of 8% (so again in-core applications could make more
progress).

thanks for the feedback Lorenzo!

Andrea