-
Notifications
You must be signed in to change notification settings - Fork 312
Best Practices
Although the master node doesn't execute any job, its functions and its sizing are crucial to the overall performance of the cluster.
When choosing the instance type to use for your master node you want to evaluate the following items:
- Cluster size: the master node orchestrates the scaling logic of the cluster and is responsible of attaching new nodes to the scheduler. If you need to scale up and down the cluster of a considerable amount of nodes then you want to give the master node some extra compute capacity.
- Shared file systems: when using shared file systems to share artefacts between compute nodes and the master node take into account that the master is the node exposing the NFS server. For this reason you want to choose an instance type with enough network bandwidth and enough dedicated EBS bandwidth to handle your workflows.
There are three hints that cover the whole range of possibilities to improve network communication.
-
Placement Group: a cluster placement group is a logical grouping of instances within a single Availability Zone. You can find more information on placement group here. You can configure the cluster to use your own placement group with
placement_group = <your_placement_group_name>or let ParalleCluster create a placement group with the "compute" strategy withplacement_group = DYNAMICMore configuration info can be found here. -
Enhanced Networking: consider to choose an instance type that supports Enhanced Networking. For more information about Enhanced Networking see here.
-
Instance bandwidth: the bandwidth scales with instance size, please consider to choose the instance type which better suits your needs, see here and here.
Increase the ulimit to allow a large number of files to be open:
# For large scale MPI runs (>1000 ranks)
echo 'sudo prlimit --pid $$ --nofile=10000:10000' >> $HOME/.bashrc
Run in a placement group:
[cluster yourcluster]
placement_group = DYNAMIC