Sampling based reporting

[nagorny]

I'm stuck trying to collect counter values over a timeline — specifically, to trace a queue size as described in the documentation.

"Using a counter in this way effectively takes the integral of X. If X were the number of entries in some queue, which was 3, 6, 1, 1, 1 over 5 cycles, this counter would be incremented by 3, then 6, then 1, 1, and finally 1 again, resulting in 12. In post-processing the delta of this type of counter (which is a discrete integral) over any range of time can be differentiated to get the average value of X over that range of time."

What kind of trigger should I use in report definition?
Even for the

trigger: 
   update-time: 1 ps

I'm getting only a small number of samples. Specifically, I'm changing the counter value every tick — 20,000 times — but I only get about 170 samples with the default clock source in the model (1 GHz, I believe).

[klingaard]

Need a little more information... in your simulation do you have multiple clocks at different frequencies? Not sure it matters, but curious. Specifically, what I'm curious about is: does 1 tick == 1 ps?

You can kinda debug this a bit but looking at the Scheduler's debug logs:

./simulator -l scheduler debug sched.out ....

WARNING: the log will get big fast. Might want to throw on a -r 10 to limit it.

What you're looking for is a trigger event being scheduled. It should be called time_trigger_event. It's scheduled every PS in your case.

[nagorny]

Hello, No in this simulation I have one clock only and it has default frequency. What I want - just print out state of several queues during simultation. I understand that it requeres a lot of storage but I'm ok with it.

What I did - I implemented some event handler where I increment some INTEGRAL counters and reschedule it recursivly for the next tick. Unfortunately I don't see all changes in the report csv file. I'm stuck with it. I think that the proper trigger in definition file should looks like

trigger:
    update-cycles: 1

But it also doesn't work for me
I experimented with update-time: 1 ps just trying to find the fatstest available sampling rate.

[klingaard]

Hmm... I'm having a tough time reproducing what you're seeing.

I took the sparta_skeleton "simulator" and had it generate a report. These are the input files:

Build the example:

cd build/example/SkeletonPipeline
make

Create the input files:

% cat > consumer_count.def 
content:
  report:
    pattern:   _global
    def_file:  simple_stats.yaml
    dest_file: consumer_counts.csv
    format:    csv
    trigger:
      update-cycles: 1
<ctrl-D>

% cat simple_stats.yaml 
content:
  subreport:
    content:
      top.consumer.stats.num_consumed: num_consumed
<ctrl-D>

Run it

./sparta_skeleton --report consumer_count.def 

Look at the output:

% wc -l consumer_counts.csv 
     103 consumer_counts.csv
% tail -n +2 consumer_counts.csv | awk '{sum += $1} END {print sum}'
100

Even works for me if I change

update-cycles: 1

to

update-time: 1 ps

[nagorny]

Thank you for examples. I tried with your def files but it doesn't work for me :(. Probably I broke something in my codebase. I'll try to recheck with original code. Sorry to bother you.

[klingaard]

What version of sparta are you using?