edgerun-galileo 0.10.4.dev1

Galileo: A framework for distributed load testing experiments

Galileo: A framework for distributed load testing experiments

This project allows users to define, run, and interact with distributed load testing experiments for distributed web-service-oriented systems. Galileo consists of two major components a user can interact with: the experiment controller shell and the galileo dashboard. The experiment controller can spawn emulated clients on workers, and control the amount of load they generate. Furthermore, a user can interact with the service routing table shell to control to which cluster node a service request is sent to.

Build

Create a new virtual environment and install all dependencies

make venv

Docker

To create a Docker image that can run galileo components, run

make docker

To create a arm32v7 Docker image that can run galileo components, run

make docker-arm

Start a worker with

cd docker/galileo-worker
docker-compose up

Compose files for arm32v7 are located in

docker/arm

Start a local dev environment, including: API, ExperimentDaemon, 1 worker, redis and database:

cd docker/dev
docker-compose up

Preparing the Example Application

We prepare the cluster with an example application. Specifically a image classification service.

Run the mxnet-model-server as a Docker container named 'mms', exposed on port 8080. For example, to start mxnet-model-server with models squeezenet and alexnet, run the following command on a cluster node:

docker run -itd --name mms -p 8080:8080 -p 8081:8081 awsdeeplearningteam/mxnet-model-server:1.0.0-mxnet-cpu mxnet-model-server --start \
--models squeezenet=https://s3.amazonaws.com/model-server/models/squeezenet_v1.1/squeezenet_v1.1.model,alexnet=https://s3.amazonaws.com/model-server/model_archive_1.0/alexnet.mar

Prepare the Experiment Worker Hosts

The devices hosting the workers that generate load need to run the experiment controller host application.

bin/run worker --logging=INFO

All runtime parameters are controlled via galileo_* environment variables. Check docker/galileo-worker/worker.env for some examples.

All environment variables, that start with galileo_, can be used as worker label when creating a client group.

I.e., if you start a worker process with the env variable galileo_zone=A, you can spawn a client group that contains only workers with this labels as follows:

g.spawn('service',worker_labels={'galileo_zone': 'A'})

Run the Experiment Controller Shell

(.venv) pi@graviton:~/edgerun/galileo $ bin/run shell
                                   __  __
 .-.,="``"=.          ____ _____ _/ (_) /__  ____
 '=/_       \        / __ `/ __ `/ / / / _ \/ __ \
  |  '=._    |      / /_/ / /_/ / / / /  __/ /_/ /
   \     `=./`.     \__, /\__,_/_/_/_/\___/\____/
    '=.__.=' `='   /____/


Welcome to the galileo shell!

Type `usage` to list available functions

galileo> usage
the galileo shell is an interactive python shell that provides the following commands

Commands:
  usage         show this message
  env           show environment variables
  pwd           show the current working directory

Functions:
  sleep         time.sleep wrapper

Objects:
  g             Galileo object that allows you to interact with the system
  show          Prints runtime information about the system to system out
  exp           Galileo experiment
  rtbl          Service routing table

Type help(<function>) or help(<object>) to learn how to use the functions.

Configure the routing table

The rtbl object in the shell allows you to set load balancing policies. Run help(rtbl) in the galileo shell. Here is an example of how to set a record for the service myservice.

galileo> rtbl.set('myservice', ['host1:8080', 'host2:8080'], [1,2])
RoutingRecord(service='myservice', hosts=['host1:8080', 'host2:8080'], weights=[1, 2])
galileo> rtbl
+---------------------------+----------------------+-----------+
| Service                   |                Hosts |   Weights |
+---------------------------+----------------------+-----------+
| myservice                 |           host1:8080 |       1.0 |
|                           |           host2:8080 |       2.0 |
+---------------------------+----------------------+-----------+

Run the Experiment Daemon

---

FIXME: THIS IS OUTDATED

---

The experiment daemon continuously reads from the blocking redis queue galileo:experiments:instructions. After receiving instructions, the controller will execute the commands and record the telemetry data published via Redis. At the end the status of the experiment will be set to 'FINISHED' and the traces, that were saved in the db by the clients, will be updated to reference the experiment.

Example Redis command to inject a new experiment (where exphost is the hostname of the experiment host):

LPUSH galileo:experiments:instructions '{"instructions": "spawn exphost squeezenet 1\nsleep 2\nrps exphost squeezenet 1\nsleep 5\nrps exphost squeezenet 0\nsleep 2\nclose exphost squeezenet"}'

you can also specify exp_id, creator, and name, otherwise some generated/standard values will be used.

You can change the database used to store the experiment data via the env galileo_expdb_driver (sqlite or mysql). To write the changes into MySQL (or MariaDB), set the following environment variables: galileo_expdb_mysql_host, galileo_expdb_mysql_port, galileo_expdb_mysql_db, galileo_expdb_mysql_user, galileo_expdb_mysql_password

You can create a mariadb docker instance with:

./bin/run-db.sh

Then execute the daemon with:

python -m galileo.cli.experimentd

Or run the script, which exports the mariadb setup from the docker container (add --logging DEBUG for output)

./bin/experimentd-mysql.sh

Run the Galileo REST API

Serve the app with gunicorn

gunicorn -w 4 --preload -b 0.0.0.0:5001 \
    -c galileo/webapp/gunicorn.conf.py \
    galileo.webapp.wsgi:api