CodeBench worker

The worker handles lifecycle of an execution job.

• Get jobs from a RabbitMQ queue
• Keep a pool of microVMs warm, and send new jobs to pre-booted VM to reduce overhead
• Run code execution jobs trough the agent
• Update back an ephemeral RabbitMQ queue for each job with status and result

The worker uses firecracker-go-sdk to communicate with the microVMs.

Requirements

• The firecracker binary in the PATH
• A rootfs in ../agent/rootfs.ext4 with the agent installed and enabled at boot
• A Linux kernel at ../../linux/vmlinux.
• CNI plugins and config (see below)

Both the rootfs and the kernel can be built with scripts in the linked repo.

CNI

In /etc/cni/conf.d/fcnet.conflist:

{
  "name": "fcnet",
  "cniVersion": "0.4.0",
  "plugins": [
    {
      "type": "ptp",
      "ipMasq": true,
      "ipam": {
        "type": "host-local",
        "subnet": "192.168.127.0/24",
        "resolvConf": "/etc/resolv.conf"
      }
    },
    {
      "type": "tc-redirect-tap"
    }
  ]
}

If you want to disable the internet access of the micro VM, you can set ipMasq to false. This will disable the IP masquerade on the host for this network.

You should also have the required CNI plugins binaries in /opt/cni/bin:

From: https://github.com/containernetworking/plugins

• host-local
• ptp

From: https://github.com/awslabs/tc-redirect-tap

• tc-redirect-tap

Demo

Start RabbitMQ:

docker-compose up -d

Start the worker:

» go run

Let's add a job to compile and run this program:

#include <stdio.h>

int main() {
   printf("Hello, World!");
   return 0;
}

We will send this job:

{
  "id": "666",
  "type": "code",
  "code": "#include <stdio.h>\r\nint main() {\r\n   printf(\"Hello, World!\");\r\n   return 0;\r\n}"
}

This is the JSON request for RabbitMQ:

{
  "routing_key": "jobs",
  "properties": {},
  "payload": "{\"id\":\"666\", \"type\": \"code\", \"code\":\"#include <stdio.h>\\r\\nint main() {\\r\\n   \\/\\/ printf() displays the string inside quotation\\r\\n   printf(\\\"Hello, World!\\\");\\r\\n   return 0;\\r\\n}\"}\r\n",
  "payload_encoding": "string"
}

Let's send it:

» curl 'http://localhost:15672/api/exchanges/%2F/amq.default/publish' -u admin:admin -d '@req.json'
{"routed":true}

The worker gets the job, gets a microVM from the pool, and handles the job:

INFO[0066]/root/codebench/worker/main.go:67 main.main() Received a message: {"id":"666", "type": "code", "code":"#include <stdio.h>\r\ni t main() {\r\n   \/\/ printf() displays the string inside quotation\r\n   printf(\"Hello, World!\");\r\n   return 0;\r\n}"}
INFO[0066]/root/codebench/worker/job.go:13 main.benchJob.run() Handling job                                  job="{666 code  #include <stdio.h>\r\nint main() {\r\n   // printf() displays the string inside quotation\r\n   printf(\"Hello, World!\");\r\n   return 0;\r\n}}"
INFO[0066]/root/codebench/worker/job_queue_rabbitmq.go:82 main.jobQueue.setjobStatus() Set job status                                status=received
INFO[0066]/root/codebench/worker/job_queue_rabbitmq.go:82 main.jobQueue.setjobStatus() Set job status                                status=running
INFO[0066]/root/codebench/worker/job.go:88 main.benchJob.run() Job execution finished                        result="{ Hello, World!}"
INFO[0066]/root/codebench/worker/job_queue_rabbitmq.go:125 main.jobQueue.setjobResult() Set job result                                jobStatus="&{666 done   Hello, World!}"
INFO[0066]/root/codebench/worker/vm.go:43 main.runningFirecracker.shutDown() stopping                                      ip=192.168.127.6
WARN[0066] firecracker exited: signal: terminated

At the same time, a new VM has been added to the missing spot in the pool.

Let's check RabbitMQ:

» docker-compose exec rabbitmq rabbitmqadmin -u admin -p admin list queues
+----------------+----------+
|      name      | messages |
+----------------+----------+
| job_status_666 | 3        |
| jobs           | 0        |
+----------------+----------+

We now have a queue corresponding to the ID of the job, with message in it:

» docker-compose exec rabbitmq rabbitmqadmin -u admin -p admin get queue=job_status_666 count=3 ackmode=ack_requeue_false
+----------------+----------+---------------+--------------------------------------------------------------------------------+---------------+------------------+-------------+
|  routing_key   | exchange | message_count |                                    payload                                     | payload_bytes | payload_encoding | redelivered |
+----------------+----------+---------------+--------------------------------------------------------------------------------+---------------+------------------+-------------+
| job_status_666 |          | 2             | {"id":"666","status":"received","stderr":"","stdout":""}          | 69            | string           | False       |
| job_status_666 |          | 1             | {"id":"666","status":"running","stderr":"","stdout":""}           | 68            | string           | False       |
| job_status_666 |          | 0             | {"id":"666","status":"done","stderr":"","stdout":"Hello, World!"} | 78            | string           | False       |
+----------------+----------+---------------+--------------------------------------------------------------------------------+---------------+------------------+-------------+

Our C program has been successfully compiled and executed!

{
  "id": "666",
  "status": "done",
  "stderr": "",
  "stdout": "Hello, World!"
}

Known issues

Cleanup when shutting down VMs does not work properly.