Kernel tested: 2.4.12-ac3+vmpatch+freeswap
Summary:
freeswap patch fixes swap deallocation issue seen in 2.4.12-ac1.
Notes:
Setting page-cluster to 2 instead of 4 helps mp3blaster in this
test, but page-cluster=2 is not as magic as in 2.4.13-pre3aa1.
page-cluster=2, 1:35 seconds of mp3 played.
page-cluster=4, 1:09 seconds of mp3 played.
Test duration was actually about 5:50, so there was a lot of
time the mp3 was not playing. It's clear to me now that mp3blaster
doesn't skip, it just doesn't play for some period of time.
Test:
Run loop of 10 iterations of Linux Test Project's "mtest01 -p80 -w"
This test attempts to allocate 80% of virtual memory and write to
each page. Simulataneously listen to mp3blaster.
Hardware:
Athlon 1333
512 Mb RAM
1024 Mb swap
page-cluster=2
Averages for 10 mtest01 runs
bytes allocated: 1255460044
User time (seconds): 2.042
System time (seconds): 4.125
Elapsed (wall clock) time: 36.066
Percent of CPU this job got: 16.60
Major (requiring I/O) page faults: 115.8
Minor (reclaiming a frame) faults: 307297.1
page-cluster=4
Averages for 10 mtest01 runs
bytes allocated: 1250846310
User time (seconds): 2.013
System time (seconds): 3.977
Elapsed (wall clock) time: 34.777
Percent of CPU this job got: 16.90
Major (requiring I/O) page faults: 109.4
Minor (reclaiming a frame) faults: 306173.0
For comparison, 2.4.12-ac1 allocated 743545241 bytes for the same test.
vmstat 1 - this is the end of one iteration, through a complete test, then
the start of a new iteration:
procs memory swap io system cpu
r b w swpd free buff cache si so bi bo in cs us sy id
2 4 1 809452 3060 1252 1172 384 4280 808 4284 371 317 14 44 43
5 2 1 892996 2432 1240 1240 132 104328 200 104308 706 335 1 20 79
1 4 2 894244 3060 1256 1240 20 32048 20 32080 387 196 8 32 60
1 4 0 9260 497880 1268 1532 636 140 952 160 1178 907 2 7 91
1 1 0 9220 496488 1288 2408 404 0 1292 0 400 633 0 0 100
2 0 0 9220 231068 1288 2428 188 0 208 0 291 608 46 53 1
2 0 1 9288 3060 1288 2324 184 64 184 72 300 585 36 64 0
0 3 2 402148 2980 1228 1448 84 74528 92 74524 462 321 1 50 49
1 1 1 502184 3320 1224 1192 292 228200 344 228236 3110 2044 5 11 84
2 2 1 502184 3064 1232 1192 20 85216 20 85232 1201 894 1 8 91
0 5 0 713996 3036 1232 820 128 254876 144 254900 3565 2101 5 16 79
1 3 0 713996 3060 1232 888 128 0 208 0 361 231 3 3 94
0 2 1 848896 3060 1232 1084 320 74940 516 74936 466 362 15 55 30
0 2 1 848896 3060 1232 1084 0 31556 0 31556 343 229 0 1 99
0 4 2 889816 3060 1248 1072 16 23316 20 23316 319 181 10 28 62
3 2 0 9208 497712 1264 1444 672 33792 1072 33832 1640 1112 2 7 91
1 1 0 9208 496600 1268 1956 512 0 1028 0 409 613 1 2 97
2 0 0 9168 301052 1288 2168 204 0 428 0 326 601 35 40 26
Conclusion:
For the mtest01/mp3blaster test , interactive performance and mp3blaster quality
is not as good as 2.4.13-pre3aa1.
For the mmap001/mp3blaster test I recently posted, 2.4.12-ac3+vmpatch,
gives better interactive performance and time than 2.4.13-pre3aa1.
--
Randy Hron