Re: [rfc][patch] select_idle_sibling() inducing bouncing on westmere

[Mike Galbraith <efault@gmx.de>]

BTW, there's something pretty darn strange going on in westmere land.
Take a peek at this, note how upset it gets at large size, and how it
then can't keep up with crusty old Q6600.  Numbers are a bit erratic,
but you'll see it.  Identical kernels, not that kernel matters. 

Q6600
marge:/usr/local/tmp/lmbench3 # !968
for m in 16K 32K 64K 128K 256K 512K 1M 2M 4M 8M; do bin/x86_64-linux-gnu/bw_tcp -P 1 -N 3 -m $m localhost; done

select_idle_sibling()              !select_idle_sibling()
0.016384 1655.83 MB/sec            1628.55 MB/sec
0.032768 1816.77 MB/sec            2086.87 MB/sec
0.065536 3148.80 MB/sec            2117.82 MB/sec
0.131072 2356.40 MB/sec            1493.96 MB/sec
0.262144 2827.00 MB/sec            1908.48 MB/sec
0.524288 3301.68 MB/sec            1908.48 MB/sec
1.048576 3359.03 MB/sec            1553.58 MB/sec
2.097152 3143.11 MB/sec            2259.11 MB/sec
4.194304 3020.53 MB/sec            1949.93 MB/sec
8.388608 2823.97 MB/sec            1868.70 MB/sec

E5620
rtbox:/usr/local/tmp/lmbench3 # !877
for m in 16K 32K 64K 128K 256K 512K 1M 2M 4M 8M; do bin/x86_64-linux-gnu/bw_tcp -P 1 -N 3 -m $m localhost; done

select_idle_sibling()              !select_idle_sibling()
0.016384 3461.28 MB/sec            2687.59 MB/sec
0.032768 4400.77 MB/sec            2939.77 MB/sec
0.065536 4517.30 MB/sec            2738.92 MB/sec
0.131072 3441.09 MB/sec            1894.25 MB/sec
0.262144 3919.65 MB/sec            2479.68 MB/sec
0.524288 1232.65 MB/sec            2341.83 MB/sec
1.048576 1230.15 MB/sec            2398.64 MB/sec
2.097152 1875.09 MB/sec            1591.16 MB/sec
4.194304 1382.21 MB/sec            1791.67 MB/sec
8.388608 1406.07 MB/sec            1078.92 MB/sec