Offload tests

Quick start

Cloud

cd x86_64
docker-compose up

Browse to $cloud_addr:1880. Import flow_remote.json.

RPI

cd armv7l
export DOCKER_HOST=$rpi_addr>:2375
docker-compose up

Point your browser to $rpi_addr:1880. Import flow_local.json. Modify the cloud-link node to point to the flow in the cloud.

Usage

Trigger the start node to test. The node injects a payload with the following properties:

• url: the url of the image to load
• batchId: a batch identifier (default: batch0)
• thresholds (optional): override the values on cloud-link node (e.g.: { "mem": 100, "cpu": 100, "temp": 80})

The workflow can be triggered by the HTTP endpoint /newjob to start a batch upload. E.g. (see submit_jobs.sh):

BATCH_ID=batch1
ADDR=resin.local:1880
IMG_URL=http://jurnsearch.files.wordpress.com/2009/07/ocr-test.jpg

for (( i = 0 ; i < 10 ; i++ )); do
  curl -X POST \
       -H"Content-type:application/json" \
       -d"{\"url\":\"$IMG_URL\",\"batchId\":\"$BATCH_ID\"}" \
       $ADDR/newjob
  sleep 60
done

Each job execution outputs an object with the times taken to execute the batch (time from the first execution of a job with a given batchId to the last execution), the process (time from when the image has been loaded to the last node) and the job (time spent in the HeavyTask flow), the offloading thresholds, the performance values by the time the job arrived and if the job was offloaded or not. The output is appended to the file /tmp/${batchId}.json. E.g.:

[
  {
    "batchId": "batch2",
    "thresholds": {
      "mem": "100",
      "cpu": "100",
      "temp": "80"
    },
    "trigger": "http",
    "times": {
      "jobstart": 1531125190896,
      "jobend": 1531125324444,
      "procstart": 1531125190815,
      "procend": 1531125324459,
      "batchstart": 1531125190815,
      "batchend": 1531125324459,
      "jobdelta": 133548,
      "procdelta": 133644,
      "batchdelta": 133644
    },
    "values": {
      "mem": 18.851716569136723,
      "cpu": 0.0384521484375,
      "temp": 53.7
    },
    "remote": 0
  },
  {
    "batchId": "batch2",
    "thresholds": {
      "mem": "100",
      "cpu": "100",
      "temp": "80"
    },
    "trigger": "http",
    "times": {
      "jobstart": 1531125200653,
      "jobend": 1531125330389,
      "procstart": 1531125200611,
      "procend": 1531125330401,
      "batchstart": 1531125190815,
      "batchend": 1531125330401,
      "jobdelta": 129736,
      "procdelta": 129790,
      "batchdelta": 139586
    },
    "values": {
      "mem": 35.41874681395163,
      "cpu": 0.071044921875,
      "temp": 60.1
    },
    "remote": 0
  }
]

The GET /count endpoint returns the number of jobs running at the moment, useful to know when the batch has finished.

The GET /monitor endpoint returns the last monitored values from the gateway.

Running an experiment

The script suite_submit_jobs.sh runs a full experiment from scratch. You just need to make sure the local workflow connects to the remote workflow and that the remote nodered is running.

It does:

• docker-compose up nodered on the gateway
• run performance-monitor
• submit the jobs (see below)
• docker-compose down
• collect the results from the gateway
• generate graphs

There are two modes of submitting modes:

• no-offloading: to check gateway behavior. It sends a total number 'n' of jobs to process, but only a number 'p' can be processed in parallel.
• offloading. It sends a total number 'n' of jobs to process, each 'p' seconds. One of {cpu, mem, temp, maxlocal} must be specified as a threshold to know when to offload.

The script uses a number of env vars, which are read from conf.sh.