Hi !
This is my first a littl more clean attemt to "measure" the precision that
rtl will reach on a system , the idea here is to get the system jitter not
in absolute terms (measurement with an osziloscope on some pin) but as
rtl sees it. it was basicaly done by having the folowing construct in a
periodic thread.
rt_task outline:
----------------
---snip---
for(vec_size=1;vec_size<=5;vec_size++){
min_diff[vec_size] = 2000000000;
max_diff[vec_size] = -2000000000;
}
for (i = 0; i < ntests; i++) {
++cnt;
pthread_wait_np();
end[0] = clock_gethrtime(CLOCK_UST);
end[1] = clock_gethrtime(CLOCK_UST);
end[2] = clock_gethrtime(CLOCK_UST);
end[3] = clock_gethrtime(CLOCK_UST);
end[4] = clock_gethrtime(CLOCK_UST);
end[5] = clock_gethrtime(CLOCK_UST);
diff[1] = end[1]-begin;
diff[2] = end[2]-end[1];
diff[3] = end[3]-end[2];
diff[4] = end[4]-end[3];
diff[5] = end[5]-end[4];
for(vec_size=1;vec_size<=5;vec_size++){
if (diff[vec_size] < min_diff[vec_size]) {
min_diff[vec_size] = diff[vec_size];
}
if (diff[vec_size] > max_diff[vec_size]) {
max_diff[vec_size] = diff[vec_size];
}
}
}
for(vec_size=1;vec_size<=5;vec_size++){
samp.run[vec_size] = min_diff[vec_size];
samp.jitt[vec_size] = (max_diff[vec_size]-min_diff[vec_size]);
}
write (fd_fifo, &samp, sizeof(samp));
---snip---
this is simply the inner loop of rtlinux/examples/measurement/rt_process.c
modified , the calls to clock_gethrtime(CLOCK_UST) are simply called
consecutive , and basicaly nothing should realy interrupt them. the inner
loop (ntest) was 500 long .
measurements were done in two cases (more fine-grain and selective measurements
will folow when I have time) , first a low-load measurement running 30 minutes
which coresponds to 3470 samples of the above thread .The system was as
unloaded as posible (see interrupts below) . The second measurement was done
at a high load-average and with a maximum of interrupt events . The limiting
factor for a longer run than 30 minutes was that I was not willing to
klick/move the mouse and hammer the keyboard for more than 30 minutes :)
low-load measurement:
---------------------
loadaverage of the linux system was between 0.00 and 0.30 .
runtime of the test 30min (approximatly) or 3470 samples (exactly).
interrupts during this lowload measurement:
-------------------------------------------
0: 172903 RTLinux virtual irq timer
1: 0 RTLinux virtual irq keyboard
2: 0 RTLinux virtual irq cascade
5: 231 RTLinux virtual irq NE2000
12: 0 RTLinux virtual irq PS/2 Mouse
13: 0 RTLinux virtual irq fpu
14: 5182 RTLinux virtual irq ide0
221: 3440 RTLinux virtual irq fifo
222: 0 RTLinux virtual irq RTL-scheduler
223: 0 RTLinux virtual irq rtl_printf
ERR: 0
I assume that the fifo interrupts where 3470 , because I had 3470 samples in
my data-file , don't know why proc only saw 3440 , and probably the other
numbers are to be seen as a good aproximation no more....
jitter between cals:
--------------------
min in all exept 2 cases was 5028 (in two casse 2514) ,
jitter here is max-min. all loops where of equal length
(500 calls)
the maximum jitter value was obviously not dependant on when the call was
issued , running five consecutive calls was done only to ensure that there
is now load effect involved (contextswitch or so).
call-number 5 4 3 2 1 jitter (ns)
17 1786 2419 912 2146 1677
6 118 88 203 122 2514
3066 1236 879 1774 959 2515
15 18 4 21 6 3352
357 307 64 549 208 3353
1 0 0 1 0 4190
1 1 1 4 2 4191
0 0 0 0 1 4192
1 0 0 0 0 5028
0 0 0 0 1 5029
0 0 2 1 3 15924
0 1 1 1 0 16762
0 1 3 0 2 16763
1 1 3 2 3 17601
0 0 0 0 1 17600
0 0 0 1 0 18438
3 0 0 2 0 18439
1 0 0 0 0 19276
0 0 1 0 0 19277
0 1 0 0 0 20114
1 0 0 0 0 20115
0 0 1 0 0 20953
0 0 0 1 0 21791
first conclusion , the maximum shown here is the smales jitter that a
rtl-thread will show , I guess there is no way to get guarantied lower
values on this hardware.
high-load measurement:
----------------------
load on the linux system was between 6.XX and 9.XX , load was produced by
multiple calls to "ls -lR /" , "ping -l 1000 broadcast-addr" and
"cp -r localdir nfs-mounted-dir" . The CPU-usage was close to 100% during
the entire test (checkt it with top -> nice 0.0% , idle 0.0%-2.0%) .
interrupts during the highload measurement:
-------------------------------------------
0: 174003 RTLinux virtual irq timer
1: 21778 RTLinux virtual irq keyboard
2: 0 RTLinux virtual irq cascade
5: 40033 RTLinux virtual irq NE2000
12: 64659 RTLinux virtual irq PS/2 Mouse
13: 0 RTLinux virtual irq fpu
14: 84042 RTLinux virtual irq ide0
221: 3502 RTLinux virtual irq fifo
222: 0 RTLinux virtual irq RTL-scheduler
223: 0 RTLinux virtual irq rtl_printf
ERR: 0
again the same holds true , the interrupt counts are not absolutly correct I
guess, because the datafile was almost the same size (3484 samples...snipped
of the 14 that where too many).
jitter between cals:
--------------------
min in most cases again was 5028 (in less than 1% min was 1677/2514/3352),
jitter here is max-min. All loops where of equal length again (500 calls).
call-number 5 4 3 2 1 jitter (ns)
16 6 10 12 11 1677
7 0 0 1 0 2514
674 13 44 4 8 2515
602 6 4 1 3 3352
1501 2831 2785 2807 2756 3353
0 1 0 3 1 4190
1 2 2 0 9 4191
0 0 0 0 2 4192
0 0 1 0 1 5028
3 3 2 0 0 5029
0 1 0 0 1 5030
0 0 1 0 0 5866
0 0 1 0 1 5867
1 0 0 0 0 12571
0 0 1 0 0 13410
1 0 0 0 0 14248
1 2 0 3 1 15086
0 1 2 1 0 15087
1 3 4 1 5 15924
0 2 4 0 3 15925
6 5 7 6 4 16762
6 8 6 5 8 16763
2 0 0 0 0 17599
13 5 11 9 9 17600
12 17 20 17 10 17601
2 0 0 0 0 18437
16 14 11 6 9 18438
28 27 31 25 25 18439
1 0 0 0 0 19275
15 10 9 13 15 19276
42 42 51 54 36 19277
1 0 0 0 0 20113
23 13 7 8 7 20114
51 69 66 72 83 20115
18 3 1 1 3 20952
72 75 89 79 79 20953
7 0 0 0 0 21790
53 64 60 55 79 21791
9 4 6 9 10 21792
11 0 0 0 0 22628
44 53 49 49 77 22629
13 11 11 21 11 22630
6 0 0 0 0 23466
50 36 36 39 69 23467
14 11 21 19 16 23468
4 0 0 0 0 24304
24 30 32 41 34 24305
19 9 23 23 20 24306
4 0 0 0 0 25142
43 22 19 24 24 25143
15 20 0 0 0 25144
3 0 0 0 0 25980
10 18 10 12 7 25981
14 19 11 10 5 25982
4 5 1 1 1 26819
5 8 1 7 0 26820
1 0 0 0 0 27657
1 1 0 3 1 27658
there is a hole between 3 and 15us which I have no explenation for , but
maby I would have to take much longer samples to realy be able to say
anything sensible about this.
interpretation attemt:
----------------------
the jitter is "Modulo 838" wich is the 8254 timmers-tick length in ns.
statistic correlation on the consecutive calls is good, event-corelation
is low to non-existent . External load clearly has an influence
comparing maxima , it shows that the interrupts or periferal
computer-components probably are NOT the cause of this jitter.
even with high load and high interrupt-rates on all connected
devices + CPU the maximum jitter "only" was increased by 27%,
although there is also a shift towards higher average values, wich
is not of so great importance because the criteria here is WORST
not AVERAGE case.
one conclusion that I am uncirtain about but I will put it here for discusion
is that since the low-load measurement shown a time of 5028ns for
clock_gethrtime(CLOCK_UST) (with the noted two exeptions of 17350 samples)
and the high-load measurement showd LOWER values in 40 cases , I assume
that the time-precision of clock_gethrtime(CLOCK_UST) is limited to something
like +/- 2*838ns (again 838 ns being a tick of the 8254). could someone
comment on this assumtion ??
a further conclusion is that rtlinux obviously is quite stable concerning
worst-case guarantee under high system/interrupt-load.
hardware:
---------
486 75MHz (genuine intel)
OS SuSE 6.2
Kernel 2.3.35
RTL V3.0.pre1
24MB-RAM
IDE-HD (Segate 4.3MB)
ISA/VESA BUS no PCI !
any suggestion on what I'm measuring here and how it could be inproved are
very wellcomed.
thx
nmg
*------------------------------------------------------------------------*
*THERE {__} Universitaet Wien *
*IS oo ) Inst. f. Materialphysik *
*LIFE O_ `_-------. Mc Guire Nicholas *
*IN (-. ,-\ *
*THE || )---< ) [EMAIL PROTECTED] *
*NET ! ||| | | privat: [EMAIL PROTECTED] *
* [EMAIL PROTECTED] *
*------------------------------------------------------------------------*
*ASCII-COW (C) 1996-2001 [EMAIL PROTECTED] *
*------------------------------------------------------------------------*
--- [rtl] ---
To unsubscribe:
echo "unsubscribe rtl" | mail [EMAIL PROTECTED] OR
echo "unsubscribe rtl <Your_email>" | mail [EMAIL PROTECTED]
----
For more information on Real-Time Linux see:
http://www.rtlinux.org/~rtlinux/
