Skip to content

Latest commit

 

History

History
55 lines (49 loc) · 2.39 KB

Watchdog.md

File metadata and controls

55 lines (49 loc) · 2.39 KB
Raw

Watchdog Tool

The watchdog component (see: toolbox/watchdog.py) allows you to monitor the conduct of results/events as the simulator executes and to end the simulation if the watched-for result/event occurs. Suppose, for instance, you were optimizing instruction throughput and wanted to know if it ever occurred that 1,000 or more cycles elapse between consecutive instruction retires in the Rangpur pipeline while executing the sum example binary:

mkdir -p /tmp/rangpur/sum
cd ${HOME}/src/nebula/pipelines/rangpur/
python3 ../../launcher.py \
    --log /tmp/rangpur/sum \
    --service ../../toolbox/stats.py:localhost:-1:-1 ../../components/simplemainmem/mainmem.py:localhost:-1:-1 \
    --config stats:output_filename:/tmp/rangpur/sum/stats.json \
    mainmem:filename:/tmp/rangpur/sum/mainmem.raw \
    mainmem:capacity:$(( 2**32 )) \
    watchdog:result_name:retire \
    watchdog:result_cycles:1000 \
    --max_instructions $(( 10**5 )) \
    -- \
    12345 \
    localhost.nebula \
    ../../examples/bin/sum 2 3 5 7 11 13

Notice the watchdog:result_name and watchdog:result_cycles parameters passed to --config. The former sets the name of the result channel occurrence to watch for; the latter sets the threshold of occurrences that trigger the end of simulation.

Similarly, if you wanted the watchdog to end simulation if ever more than 1,000 cycles elapse between consecutive instruction commits on the Bergamot pipeline:

mkdir -p /tmp/bergamot/sum
cd ${HOME}/src/nebula/pipelines/bergamot/
python3 ../../launcher.py \
    --log /tmp/bergamot/sum \
    --service ../../toolbox/stats.py:localhost:-1:-1 ../../components/simplemainmem/mainmem.py:localhost:-1:-1 \
    --config stats:output_filename:/tmp/bergamot/sum/stats.json \
    mainmem:filename:/tmp/bergamot/sum/mainmem.raw \
    mainmem:capacity:$(( 2**32 )) \
    watchdog:event_name:commit \
    watchdog:event_cycles:1000 \
    --max_instructions $(( 10**5 )) \
    -- \
    12345 \
    localhost.nebula \
    ../../examples/bin/sum 2 3 5 7 11 13

Notice that watchdog:event_name and watchdog:event_cycles are used (rather than watchdog:result_name and watchdog:result_cycles), since finished instructions are signaled on the event channel in the Bergamot pipeline model (rather than the results channel).