For your ML model training tasks you have to provision an appropriate compute resource.
Training runs are executed periodically and in idle periods you don't need the resource, but during training runs, the compute has to cope with really high loads.
How do you fulfil the requirement while avoiding excess cost?
Click on the arrows to vote for the correct answer
A. B. C. D.Answer: C.
Option A is incorrect because it is a single VM which doesn't scale down when idle and doesn't scale up with peak demand; it has to be managed manually which induces a risk of unexpected costs.
Option B is incorrect because remote VMs are not managed by Azure, scale up/down must be separately managed, which might not fulfil the requirements of peaks in training runs.
Option C is CORRECT becauseAzure compute clusters are the most flexible and cost-effective environments for train experiments.
Clusters, as a multi-node environment can scale up for training runs and autoscale to 0 nodes when idle.
Option D is incorrect because AKS is a great and powerful compute environment inference runs but it is unnecessarily expensive, therefore not recommended for training/development purposes.
Use Compute Clusters instead.
Reference:
To fulfill the requirement of training an ML model on a compute resource that can handle high loads while avoiding excess cost during idle periods, there are several options available in Azure:
A. Provision a managed Azure Compute Instance: An Azure Compute Instance is a managed virtual machine that is pre-configured with popular data science tools and frameworks. It provides a scalable and cost-effective solution for running data science workloads. With this option, you can provision the compute resource only when you need it, and deprovision it when it's not in use. This can help to reduce costs, as you only pay for the compute resource while it's being used.
B. Attach a remote VM as a compute: Another option is to use an existing remote VM as a compute resource for training your ML model. This can be a cost-effective solution if you already have a VM that meets the requirements for your training workload. You can also turn off the VM when it's not in use to save costs.
C. Provision Azure ML Compute Cluster: Azure ML Compute Cluster is a managed compute resource that is designed for running training workloads in Azure Machine Learning. It can automatically scale up or down based on the workload and provides a cost-effective solution for running training jobs. You can specify the maximum number of nodes to be used in the cluster, and the cluster will automatically adjust the number of nodes based on the workload.
D. Make use of Azure Kubernetes Service: Azure Kubernetes Service (AKS) provides a managed Kubernetes cluster for deploying containerized applications. You can use AKS to run your ML training workload in a containerized environment. This can provide a scalable and cost-effective solution for running training workloads, as you can easily scale up or down the number of nodes in the cluster based on the workload.
In summary, all of the above options can be used to fulfill the requirement of training an ML model on a compute resource that can handle high loads while avoiding excess cost. The best option will depend on the specific requirements of the workload, the available resources, and the budget.