YYYYYYY YYYYYYY AAA CCCCCCCCCCCCC SSSSSSSSSSSSSSS
Y:::::Y Y:::::Y A:::A CCC::::::::::::C SS:::::::::::::::S
Y:::::Y Y:::::Y A:::::A CC:::::::::::::::CS:::::SSSSSS::::::S
Y::::::Y Y::::::Y A:::::::A C:::::CCCCCCCC::::CS:::::S SSSSSSS
YYY:::::Y Y:::::YYY A:::::::::A C:::::C CCCCCCS:::::S
Y:::::Y Y:::::Y A:::::A:::::A C:::::C S:::::S
Y:::::Y:::::Y A:::::A A:::::A C:::::C S::::SSSS
Y:::::::::Y A:::::A A:::::A C:::::C SS::::::SSSSS
Y:::::::Y A:::::A A:::::A C:::::C SSS::::::::SS
Y:::::Y A:::::AAAAAAAAA:::::A C:::::C SSSSSS::::S
Y:::::Y A:::::::::::::::::::::AC:::::C S:::::S
Y:::::Y A:::::AAAAAAAAAAAAA:::::AC:::::C CCCCCC S:::::S
Y:::::Y A:::::A A:::::AC:::::CCCCCCCC::::CSSSSSSS S:::::S
YYYY:::::YYYY A:::::A A:::::ACC:::::::::::::::CS::::::SSSSSS:::::S
Y:::::::::::Y A:::::A A:::::A CCC::::::::::::CS:::::::::::::::SS
YYYYYYYYYYYYYAAAAAAA AAAAAAA CCCCCCCCCCCCC SSSSSSSSSSSSSSS
master.py
andmaster_utils
worker.py
config_utils/
: utilities related to reading theconfig.json
file which defines workers.job_utils/
: definitions of tasks, jobs and messages shared by master and workers.extras/
: extra scripts such as config generator, sending requests to master, etc.start_workers.py
: starts all workers defined inconfig.json
locally with the correct ports. Workers are stopped once the script exits.
Documentation - report and architecture
-
Edit the
config.json
file with correct worker details. Or generate one. -
First the master
python3 master.py # uses the random scheduler
OR
python3 master.py rr # uses the round robin scheduker
OR
python3 master.py ll # uses the least loaded scheduler
-
Now the worker
python3 start_workers.py # uses the config.json
-
Now run
extras/requests.py
with the number of requests you want to send to master node. It should send and finish. Let it run in the background. Example:python3 extras/requests.py 100
For more customized testing,
requests_eval.py
python3 extras/requests_eval.py # custom task generated
Copy over the required logs to analysis/
.
cd analysis
python3 yacs_logs_analysis.py # task 1
python3 graph_task2.py # task 2
The external-workers
branch has experimental support for non-local workers. Tested on a Raspberry Pi node.