[ltt-dev] Experiments on CPU time statistics

[ltt-dev] Experiments on CPU time statistics

[Francis Giraldeau]

Hi,

I'm doing some experiments with CPU statistics, and I think that those  
computed by lttv are wrong. I did 3 tests, and for each command the  
time utility was used to get a baseline, and the command has been run  
between lttctl -C and lttctl -D to get a tight window around the test  
process.

=================================================
trace-bonnie.sh : start bonnie++ process for 10s

time statistics:
real	0m10.715s

user	0m0.420s

sys	0m6.350s


lttv statistics:

events count        : 11088588

cpu time            : 80.051904112

elapsed time?       : 544.593369365

cumulative cpu time : 6.747053976

=================================================
sleep 1s : start a sleep of 1s

time statistics:
real	0m1.002s

user	0m0.000s

sys	0m0.000s


lttv statistics:
events count        : 94929


cpu time            : 7.841686487

elapsed time        : 34.566952675

cumulative cpu time : 0.036465023


=================================================
trace-burn.sh 1s * 8 process simultan?s

time statistics:
real	0m1.031s

user	0m7.650s

sys	0m0.020s


lttv statistics:
events count        : 95402

cpu time            : 8.584145675

elapsed time        : 33.845893813

cumulative cpu time : 0.037255816
=================================================

I was expecting lttv enlapsed time to be close to real time, lttv  
cumulative cpu time to be greater or equal to lttv cpu time, and lttc  
cpu time to match user + sys time. Does it make sens, or maybe I'm  
getting something wrong?

Cheer,

Francis

[ltt-dev] Experiments on CPU time statistics

[Mathieu Desnoyers]

* Francis Giraldeau (Francis.Giraldeau at USherbrooke.ca) wrote:
> Hi,
>
> I'm doing some experiments with CPU statistics, and I think that those  
> computed by lttv are wrong. I did 3 tests, and for each command the time 
> utility was used to get a baseline, and the command has been run between 
> lttctl -C and lttctl -D to get a tight window around the test process.
>
> =================================================
> trace-bonnie.sh : start bonnie++ process for 10s
>
> time statistics:
> real	0m10.715s
>
> user	0m0.420s
>
> sys	0m6.350s
>
>
> lttv statistics:
>
> events count        : 11088588
>
> cpu time            : 80.051904112

cpu time should be good. How many cores does you machine have ? 8 ?

>
> elapsed time?       : 544.593369365

Elapsed time is the total of the per-thread elapsed time I think (it's
really been a while since I played with this code though). It would make
sense if you have 54 threads in your system.

I don't have entire confidence that the elapsed time is correct.

>
> cumulative cpu time : 6.747053976

cumulative cpu time is known to be wrong.

Thanks,

Mathieu

>
> =================================================
> sleep 1s : start a sleep of 1s
>
> time statistics:
> real	0m1.002s
>
> user	0m0.000s
>
> sys	0m0.000s
>
>
> lttv statistics:
> events count        : 94929
>
>
> cpu time            : 7.841686487
>
> elapsed time        : 34.566952675
>
> cumulative cpu time : 0.036465023
>
>
> =================================================
> trace-burn.sh 1s * 8 process simultan?s
>
> time statistics:
> real	0m1.031s
>
> user	0m7.650s
>
> sys	0m0.020s
>
>
> lttv statistics:
> events count        : 95402
>
> cpu time            : 8.584145675
>
> elapsed time        : 33.845893813
>
> cumulative cpu time : 0.037255816
> =================================================
>
> I was expecting lttv enlapsed time to be close to real time, lttv  
> cumulative cpu time to be greater or equal to lttv cpu time, and lttc  
> cpu time to match user + sys time. Does it make sens, or maybe I'm  
> getting something wrong?
>
> Cheer,
>
> Francis
>
>
> _______________________________________________
> ltt-dev mailing list
> ltt-dev at lists.casi.polymtl.ca
> http://lists.casi.polymtl.ca/cgi-bin/mailman/listinfo/ltt-dev

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com

[ltt-dev] Experiments on CPU time statistics

[Michel Dagenais]

>> time statistics:
>> real	0m10.715s
>>
>> user	0m0.420s
>>
>> sys	0m6.350s
...
> Elapsed time is the total of the per-thread elapsed time I think (it's
> really been a while since I played with this code though). It would make
> sense if you have 54 threads in your system.
>
> I don't have entire confidence that the elapsed time is correct.

As a side note, while the elapsed time provided by "time" is probably 
reliable, the user and system time are obtained by sampling under 
asumptions which are sometimes not met. Arbitrarily large deviations 
from the reality are possible. The CPU time computed from the LTTng 
trace data is exact, not statistical.

[ltt-dev] Experiments on CPU time statistics

[Francis Giraldeau]

Mathieu Desnoyers <compudj at krystal.dyndns.org> a ?crit?:
>> =================================================
>> trace-bonnie.sh : start bonnie++ process for 10s
>>
>> time statistics:
>> real	0m10.715s
>> user	0m0.420s
>> sys	0m6.350s
>>
>>
>> lttv statistics:
>> events count        : 11088588
>> cpu time            : 80.051904112
>
> cpu time should be good. How many cores does you machine have ? 8 ?

Thanks for explainations. Yes, the system has 8 cores. The maximum CPU  
time for 1 second is hence 8 seconds.

>> =================================================
>> sleep 1s : start a sleep of 1s
>>
>> time statistics:
>> real	0m1.002s
>> user	0m0.000s
>> sys	0m0.000s
>>
>>
>> lttv statistics:
>> cpu time            : 7.841686487

How then can we explain that on an idle system, the CPU time is yet  
(almost) the number of CPUs times the trace duration?

Cheer,

Francis