Re: 2.6.13-rt6, ktimer subsystem

[Thomas Gleixner <tglx@linutronix.de>]

On Wed, 2005-09-14 at 12:38 -0700, George Anzinger wrote:

> Well, having spent a bit of time looking at the code it appears that a 
> lot of the ideas we looked at and discarded (see 
> high-res-timers-discourse@lists.sourceforge.net) are in this.  Shame it 
> was all done with out reference or comment to that list, anyone on it or 
> even the lkml.

Well, I'm considering to wear sackcloth and ashes. But this seems like
the pot calling the kettle back as I don't remember a single relevant
reference/comment from you on the UTIME/KURT mailing list after your
UTIME->HRT fork.

> A few of the top issues:
> 
> time in nanoseconds 64-bits, requires a divide to do much of anything 
> with it.  Divides are slow and should be avoided if possible.  

The divides are rare and definitely not in the hot pathes. I'm sure that
they can be replaced by some intellegent scaled math algorithm if it
turns out to be necessary. The hot path instructions are simple
add/sub/cmp which are less/equal expensive on a 32bit machine to an
operation on struct timespec or an jiffies/arch_cycles pair. The non
nsec based implementation gives a burden to 64bit machines and is
provable wrong in the aspect of summing rounding errors of interval
timers. 

> This is especially true in the embedded market.

I'm well aware of the embedded market constraints.


> The rbtree is a high overhead tree.  I suspect performance problems 
> here.  

1. rbtree is available out of the box

2. rbtree is proven to be efficient - at least there are a couple of
performance relevant users relying on it e.g. mm, ext3

3. I did insertion/removal tests with 10k entries (<2us on a 1GHz box in
userspace). This is way below the experienced and reproducible trouble
of recascading. The penalty is completely on the thread which owns the
timer for non signal related functions. For signal related functions
only the removal on expiry is a penalty for the complete system (in the
softirq)

The cascading is a penalty for the complete system all the time.

Performance is a strawman argument here. You know very well that > 90%
of the timers are inaccurate "timeout" timers related to I/O,
networking, devices. Most of those never expire (the positive feedback
removes the timer before expiry) and those timers have no constraint to
be accurate, except for the fact that they have to detect an
device/network problem at some time. In this case it is completely
irrelevant whether the timeout occures n msecs earlier or later.

> If it is the right answer here, then why not use it for normal 
> timers?  A list of timers is a rather unique thing and, I think, 
> deserves a management structure that accounts for the fact that the 
> elements in the tree are perishable.

The first goal was to seperate out the timers from the timeout API and I
believe that this seperation is necessary. 

The implementation of ktimers is not at all restricted to the timers we
addressed for now, it can also be utilized for the timeout API without
much effort. 

The performance improvement is enourmous despite the alleged 64bit math
overhead.

Testcase on a 600MHz CeleronM, 512MB RAM:

16 cyclic SCHED_FIFO tasks using clock_nanosleep(ABSTIME)
base interval = 1000 us, offset per task = 50 us 
tracing each latency value to disk

16 cyclic SCHED_OTHER tasks using clock_nanosleep(ABSTIME)
base interval = 1000 us, offset per task = 500 us
tracing each latency value to disk

while true; do hackbench 50; done
cat /dev/hda | nc atlas 4711

This injects a (insane) load avg of >600 permantely

2.6.13-rt4 (cascade based implementation)
16 cyclic SCHED_FIFO tasks using clock_nanosleep(ABSTIME)

task	loops		min	max	average	sigma	prio
0	      999999	5	869	19	20	80
1	      999999	9	883	18	22	79
2	      999999	9	927	19	23	78
3	      999999	5	908	21	28	77
4	      999999	0	1056	22	33	76
5	      999999	0	973	23	33	75
6	      999999	0	926	23	33	74
7	      999999	1	893	24	33	73
8	      999999	2	942	23	34	72
9	      999999	1	868	24	34	71
10	      999999	0	912	23	34	70
11	      999999	0	911	28	46	69
12	      999999	0	912	28	46	68
13	      999999	9	967	28	46	67
14	      999999	9	954	28	46	66
15	      999999	9	946	28	46	65


2.6.13-rt11 (ktimers based implementation)
16 cyclic SCHED_FIFO tasks using clock_nanosleep(ABSTIME)

task	loops		min	max	average	sigma	prio
0	      999999	9	76	20	4	80
1	      999999	8	84	20	4	79
2	      999999	8	98	21	4	78
3	      999999	9	121	20	4	77
4	      999999	8	124	20	4	76
5	      999999	9	140	20	4	75
6	      999999	10	103	22	5	74
7	      999999	9	99	21	5	73
8	      999999	8	95	21	5	72
9	      999999	9	148	21	5	71
10	      999999	9	141	22	6	70
11	      999999	9	143	22	5	69
12	      999999	8	129	20	4	68
13	      999999	9	149	21	5	67
14	      999999	9	135	21	5	66
15	      999999	8	230	22	5	65


> It appears that the "monotonic_clock" is being used to drive ktimers. 
> The "monotonic_clock" was NEVER meant to poke outside of the kernel.  It 
> is a raw kernel clock that is only required to be monotonic with nothing 
> said about accuracy.  It should NOT be confused with CLOCK_MONOTONIC 
> which is directly tied to xtime and therefor is ntp corrected.

I'm well aware of that and this is a completely different playground. 

The ktimers implementation is _independent_ of the clock sources. The
current clock source implementation may suck, but thats not a problem of
ktimers at all. Its a problem of arch/XXX/timeYY and kernel/time.c. I
did not spend too much time on that as John Stultz is working on the
timesource and timeofday stuff and I really hope that this gets
accepted.

The relation ship of clock sources and their accuracy is also a
worthwhile field of discussion.

> These are only the concerns I have from having a rather quick look at 
> the code.  I am sure that there are other issues...

It would be hubristic to deny that :)

OTH, 

- The posix timer tests run all successful, except the broken 2timertest
which fails on any other HRT kernel too and the sleep to long for real
timers when the clock is set backwards, which is easily solvable
(working on that).

- The performance improvements in combination with simpler code is an
argument of itself


tglx