2.5.42-mm2 contest results

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Here are the surprisingly different results from 2.5.42-mm2 with the contest benchmark (http://contest.kolivas.net). This was run with pagetable sharing enabled. Older results hidden for clarity.

noload:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.41-mm3 [1]          74.4    93      0       0       1.11
2.5.42 [2]              72.5    93      0       0       1.08
2.5.42-mm2 [3]          79.0    92      0       0       1.18

Didn't believe it the first time so I ran it twice more and ran 2.5.42 again to make sure something didn't change on my machine, but definitely this was slower than 2.5.42. When the kernel compile starts on the flushed ram machine with no background load a lot more disk activity seems to occur for the first five or so seconds compared to other kernels. I'm not sure if this is related to the way the memory and swap is flushed prior to the test or just this kernel. This may have affected all the following results too.

process_load:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.41-mm3 [1]          95.5    75      31      28      1.42
2.5.42 [1]              98.0    69      44      33      1.46
2.5.42-mm2 [2]          104.5   72      31      30      1.56

Slower again, without an increase in the loads.

ctar_load:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.41-mm3 [1]          92.1    81      1       5       1.37
2.5.42 [1]              96.7    80      1       7       1.44
2.5.42-mm2 [2]          102.3   79      1       6       1.52

Slower; lack resolution of number of loads makes it difficult to determine if it's significant. Same with a few of the other loads below.

xtar_load:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.41-mm3 [1]          215.9   34      3       7       3.21
2.5.42 [1]              112.7   66      1       7       1.68
2.5.42-mm2 [2]          195.0   41      2       6       2.90

Close to 2.5.41-mm3. Not enough runs to show if it is significant

io_load:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.41-mm3 [1]          312.4   25      20      11      4.65
2.5.42 [1]              849.1   9       69      12      12.64
2.5.42-mm2 [2]          250.7   34      15      10      3.73

This is showing an improvement with some of the better io load results shown by a 2.5 kernel.

read_load:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.41-mm3 [1]          102.0   74      6       4       1.52
2.5.42 [1]              102.0   75      8       5       1.52
2.5.42-mm2 [2]          109.0   75      7       4       1.62

list_load:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.41-mm3 [1]          95.9    74      1       22      1.43
2.5.42 [1]              97.5    71      1       20      1.45
2.5.42-mm2 [2]          105.3   72      1       24      1.57

mem_load:

Kernel [runs]           Time    CPU%    Loads   LCPU%   Ratio
2.5.41-mm3 [2]          107.1   68      27      2       1.59
2.5.42 [1]              104.0   72      30      3       1.55
2.5.42-mm2 [2]          121.2   65      30      2       1.80

Con
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.0.7 (GNU/Linux)

iD8DBQE9qVizF6dfvkL3i1gRAm+/AJ9RTjhAPz+YeDa4kNyLgR2t3b8prACfQoAk tb39kuDH4F9N7ROqwWl6RHU=
=/aJT
-----END PGP SIGNATURE-----

Well something is burning CPU there, and I do not know what it is. Things are normal here:

With 2.5.42-mm2++, shared pagetables enabled:
make -j6 bzImage 416.97s user 33.94s system 374% cpu 2:00.28 total

With 2.5.42+last night's BK:
make -j6 bzImage 416.09s user 33.04s system 370% cpu 2:01.15 total

With 2.5.42:
make -j6 bzImage 416.56s user 32.49s system 375% cpu 1:59.69 total

So. Could you please profile the `noload' run? Just enable oprofile and local IO APIC and do:

- add "idle=poll" to your kernel boot commandline
- sudo rm -rf /var/lib/oprofile
- sudo op_start --vmlinux=/boot/vmlinux --ctr0-event=CPU_CLK_UNHALTED \
--ctr0-count=600000
- time make -j4 bzImage ; sudo op_stop
- oprofpp -l -i /boot/vmlinux
- sudo killall oprofiled

(Do the `make' quickly after the op_start to avoid bogus idle time).

Thanks.