Policies

TL;DR

Use Job to run batch jobs and set right resources request; use Deployment if you need a long-running pod and set minimal resources request. Use Pod otherwise, it will be destroyed in 6 hrs.

Use monitoring to set the requests right.

Avoid wasting resources: if you’ve requested something, use it, and free up once computation is done.

Namespace

One of the main concepts of running workloads in kubernetes is a namespace. A namespace creates an isolated environment in which you can run your pods. It is possible to invite other users to your namespace and it is possible to have access to multiple namespaces.

Pod and Container

A Kubernetes pod is a group of containers that are deployed together on the same host. If you frequently deploy single-container pods, you can generally replace the word “pod” with “container”.

Memory allocation

Kubernetes scheduler tries to accommodate all pods on the set of nodes according to pods definitions and node capacities. To help it serve you the best way, you need to define in your pod the resources you’re going to consume. This is done using the Resource Limits and Requests section.

A request is what will be reserved for your pod on a node for scheduling purposes. A limit is the maximum which your pod should never exceed. If pod goes over its memory limit, it WILL BE KILLED. If your pod was suddenly killed, please make sure you’ve got the limits set up right.

While it’s important to set the Limit properly, it’s also important to not set the Request too high. Your request should be as close as possible to the average resources you’re going to consume, and limit should be a little higher than the highest peak you’re expecting to have. Use monitoring to set your requests and limits right.

Interactive use vs batch

There are so called operators to control the behaviour of pods. Since pods don’t stop themselves in normal conditions, and don’t recover in case of node failure, we assume every pod running in the system without any controller to be interactive – started for a short period of time for active development / debugging. We limit those to request a maximum of 2 GPUs, 8 GB RAM and 2 CPU cores. Such pods will be destroyed in 6 hours, unless you request an exception for your namespace (in case you run jupyterhub or some other application controlling the pods for you).

If you need to run a larger and longer computation, you need to use one of available Workload Controllers. We recommend running those as Jobs - this will closely watch your workload and make sure it ran to completion (exited with 0 status), shut it down to free up the resources, and restart if node was rebooted or something else has happened. Please see the guide on using those. You can use Guaranteed QoS for those.

In case you need some pod to run idle for a long time, you can use the Deployment controller. Make sure you set minimal request and proper limits for those to get the Burstable QoS

Requesting GPUs

When you request GPUs for your pod, nobody else can use those until you stop your pod. You should only schedule GPUs that you can actually use. The only reason to request more than a single GPU is when your GPU utilization is close to 100% and you can leverage more.

GPUs are a limited resource shared by many users. If you plan on deploying large jobs (>100 GPUs) please present a plan in rocketchat

Workloads purging (2 weeks max runtime)

We currently have more than a thousand users in the system, and many of them leave behind their deployments when their computation is done. There’s no way for us to know whether some pod is useful or is abandoned. To clear the abandoned deployments, there’s periodic process which destroys the workloads created more than 2 weeks ago. In case you’re running some permanent service and would want us to keep it running, you can contact admins in RocketChat and ask for an exception. Please provide an estimated period of service functioning and brief description of what the service does. Login pods can’t be added to exceptions list, since those are considered temporary and we need to be sure those are cleaned when not needed.

For workloads not in exceptions list you will get 3 notifications after which your workload will be deleted. Any data in persistent volumes will remain.

System load

While you’re welcome to use as much resources as needed, using those inefficiently is causing problems to others. The CPU load of the system consists of user load, system load and several others. If you run top, you’ll see something like:

Cpu(s): 26.9%us, 1.5%sy, 0.0%ni, 71.5%id, 0.0%wa, 0.0%hi, 0.1%si, 0.0%st

Which means 27.9% of all system CPU time is spent on user tasks (computing), 1.5% on system stuff (kernel tasks), 71.5 idle, and so on. If system (kernel) load is more than ~15%, it indicates a problem with users code. Usually it’s caused by too many threads (system is spending too much time just switching between those), or inefficient file I/O (lots of small files processed by too many workers) or something similar. In this case the system time is wasted not on processing.

If you were told your pod is causing too much load, look at your code and try to figure out where it’s spending too much kernel time instead of computations.