Better resolution using the hrtimers infrastructure

[John <me@privacy.net>]

Hello,

I've been experimenting with the high-resolution timer subsystem on the
x86 platform (specifically, P4 2.8 GHz) running Linux 2.6.16.28. (LAPIC
and IOAPIC turned on, pre-emptible kernel, HZ=250)

I wrote a small app to create a POSIX timer (timer_create(),
timer_settime(), etc) that fires with a given period. The scheduling
policy is set to SCHED_RR. After some time, the process writes the
distribution of "elapsed time between signals" to a file, and exits. I
then post-process this file to output average time between signals,
standard deviation, occurences +/- 5 µs and occurences +/- 10 µs.

I first compiled a kernel with a resolution of 5000 ns.
I ran 10 concurrent tasks with T = 500+13*i µs
RUNTIME = 40000 s

AVG=500.0000119875 STDDEV=4.7221250605
P(495 <= L <= 505) = 0.8177218875
P(490 <= L <= 510) = 0.9630963875

AVG=512.9983118837 STDDEV=4.6444950150
P(508 <= L <= 518) = 0.8248110095
P(503 <= L <= 523) = 0.9648220375

AVG=525.9978541435 STDDEV=4.5872769426
P(521 <= L <= 531) = 0.8294680508
P(516 <= L <= 536) = 0.9662504459

AVG=538.9976678278 STDDEV=4.6550001682
P(534 <= L <= 544) = 0.8233726222
P(529 <= L <= 549) = 0.9647691136

AVG=551.9962816176 STDDEV=4.6202416645
P(547 <= L <= 557) = 0.8261095807
P(542 <= L <= 562) = 0.9654630987

AVG=564.9992072061 STDDEV=4.7382868551
P(560 <= L <= 570) = 0.8201386764
P(555 <= L <= 575) = 0.9631647119

AVG=577.9969566710 STDDEV=4.6799294742
P(573 <= L <= 583) = 0.8199662908
P(568 <= L <= 588) = 0.9637724766

AVG=591.0001112558 STDDEV=4.6366848287
P(586 <= L <= 596) = 0.8273179260
P(581 <= L <= 601) = 0.9651421699

AVG=604.0009058943 STDDEV=4.7022062409
P(599 <= L <= 609) = 0.8253007146
P(594 <= L <= 614) = 0.9627097645

AVG=617.0019377040 STDDEV=4.7598557221
P(612 <= L <= 622) = 0.8164446084
P(607 <= L <= 627) = 0.9616794710


I then recompiled my kernel with a resolution of 800 ns.
I ran 10 concurrent tasks with T = 600+13*i µs
RUNTIME = 50000 s

AVG=599.9986703520 STDDEV=4.3792286239
P(595 <= L <= 605) = 0.8466944914
P(590 <= L <= 610) = 0.9707216439

AVG=612.9981400476 STDDEV=4.1769930447
P(608 <= L <= 618) = 0.8585605458
P(603 <= L <= 623) = 0.9756196900

AVG=625.9983042536 STDDEV=4.2016958416
P(621 <= L <= 631) = 0.8569019831
P(616 <= L <= 636) = 0.9753284384

AVG=638.9985581093 STDDEV=4.2072766644
P(634 <= L <= 644) = 0.8532895995
P(629 <= L <= 649) = 0.9746377576

AVG=651.9989919428 STDDEV=4.1936623871
P(647 <= L <= 657) = 0.8576286791
P(642 <= L <= 662) = 0.9757251792

AVG=664.9976003076 STDDEV=4.2224641203
P(660 <= L <= 670) = 0.8541156365
P(655 <= L <= 675) = 0.9748352293

AVG=677.9992813471 STDDEV=4.2120672530
P(673 <= L <= 683) = 0.8598737385
P(668 <= L <= 688) = 0.9752509658

AVG=690.9987525239 STDDEV=4.2379530226
P(686 <= L <= 696) = 0.8548963154
P(681 <= L <= 701) = 0.9744069907

AVG=703.9993729190 STDDEV=4.3323158412
P(699 <= L <= 709) = 0.8449093344
P(694 <= L <= 714) = 0.9731234341

AVG=717.0005565641 STDDEV=4.2230279084
P(712 <= L <= 722) = 0.8579009228
P(707 <= L <= 727) = 0.9744286679

The resolution (STDDEV) was slightly better, but I un-scientifically
changed two parameters. (I'll have to run more tests.)


Finally, I ran only one task with T = 600 µs

AVG=600.0036125280 STDDEV=1.6630613701
P(595 <= L <= 605) = 0.9997573360
P(590 <= L <= 610) = 0.9998031160


The resolution is markedly better (unsurprisingly).

Is it possible to improve the resolution? Would an HPET-enabled
motherboard help? The source code seems to suggest the time stamp
counter is used over any HPET. I've read that IBM provides high-quality
HPET, whereas some southbridge manufacturers barely meet the Intel spec.
Would other hardware (PCI board? PCI-Express board?) help in improving
the resolution?

Thanks for reading this far.

Regards.