big rt latencies compared to non rt

big rt latencies compared to non rt

[Angelo Dureghello]

Hi all,

i am running some RT tests on imx8qxp.
Kernel is a imx (nxp) modified kernel, but with
rt patch (5.4.70) that at least applies properly.

Linux version 5.4.70-rt40-00120-g3e7dde1168e5-dirty

I am running some histograms by

cyclictest --duration=180 -t=1 -m -Sp99 \
         -i2000 -h3000 -q > histogram

Average is near 16uS but there are some outside
values near 40000, as

000009 000000    000000    000000    000000
000010 000001    000000    000002    000000
000011 000002    000000    000000    000000
000012 000023    000006    000008    000004
000013 001565    001504    002010    001367
000014 000425    001959    000649    000700
000015 001447    001917    001129    001415
000016 034084    026118    040328    032334
000017 040537    015000    033773    041092  <-----
000018 005771    009092    007999    007512
000019 002709    023378    001790    003117
000020 000167    006238    000109    000136
000021 000135    002369    000099    000093
000022 000039    000622    000025    000041
000023 000153    000123    000026    000023
000024 000986    000178    000564    000500
000025 001872    000292    001291    001515
000026 000049    000269    000112    000081
000027 000010    000561    000011    000014
000028 000003    000296    000016    000009
000029 000005    000035    000014    000002
000030 000001    000008    000006    000000
000031 000001    000008    000006    000001
000032 000011    000002    000004    000000


Rest of the histogram is mostly 00000 values.

As far as you may know, is there something i am
missing for this cpu ?

Thanks a lot,

Regards,
angelo

Re: big rt latencies compared to non rt

[Jonathan Schwender]

Hi Angelo,

>
> Average is near 16uS but there are some outside
> values near 40000, as
You are reading the table wrong. The leftmost column is the microsecond 
bucket. The other columns give the amount of times a latency was inside 
that microsecond bucket for each measurement thread.
> 000017 040537    015000    033773    041092  <-----

For example, in this line CPU 0 had 40537 latencies in the 17 us bucket, 
while CPU 1 had 15000 measured latencies of 17 us (and so forth).

You should also check the line at the End `# Max Latencies: xxxxx`, 
which will give you the maximum latency, and the line below which shows 
you how often the histogram overflowed.

> Thanks a lot,
>
> Regards,
> angelo
>
Hope I could help,


Jonathan

Re: big rt latencies compared to non rt

[John Ogness]

Hi Angelo,

On 2021-05-11, Angelo Dureghello <angelo@kernel-space.org> wrote:
> cyclictest --duration=180 -t=1 -m -Sp99 \
>          -i2000 -h3000 -q > histogram
>
> Average is near 16uS but there are some outside
> values near 40000, as

The first column is the latency in microseconds. The other columns
(2nd-5th) is the count of times that latency was seen for a CPU.

> 000009 000000    000000    000000    000000
> 000010 000001    000000    000002    000000
> 000011 000002    000000    000000    000000
> 000012 000023    000006    000008    000004
> 000013 001565    001504    002010    001367
> 000014 000425    001959    000649    000700
> 000015 001447    001917    001129    001415
> 000016 034084    026118    040328    032334
> 000017 040537    015000    033773    041092  <-----

CPU3 saw a latency of 17us 41092 times.

John Ogness

Re: big rt latencies compared to non rt

[Angelo Dureghello]

Hi John and Jonathan,

thanks a lot for the help.

On 11/05/21 10:14 PM, John Ogness wrote:
> Hi Angelo,
> 
> On 2021-05-11, Angelo Dureghello <angelo@kernel-space.org> wrote:
>> cyclictest --duration=180 -t=1 -m -Sp99 \
>>           -i2000 -h3000 -q > histogram
>>
>> Average is near 16uS but there are some outside
>> values near 40000, as
> 
> The first column is the latency in microseconds. The other columns
> (2nd-5th) is the count of times that latency was seen for a CPU.
> 
>> 000009 000000    000000    000000    000000
>> 000010 000001    000000    000002    000000
>> 000011 000002    000000    000000    000000
>> 000012 000023    000006    000008    000004
>> 000013 001565    001504    002010    001367
>> 000014 000425    001959    000649    000700
>> 000015 001447    001917    001129    001415
>> 000016 034084    026118    040328    032334
>> 000017 040537    015000    033773    041092  <-----
> 
> CPU3 saw a latency of 17us 41092 times.
> 
> John Ogness
> 

-- 
Angelo Dureghello