Dynamic partition Working principles

Physical machines are members of a Batch System cluster (LSF) and a Cloud (Openstack) cluster at the same time; furthermore, they can only be active in a mutually exclusive manner on one and one only cluster at any time. This implicitly defines a partition of hosts active at batch or cloud side.

We refer to active Hosts in the batch partition as "Worker Nodes", and to active hosts in the cloud partition as "Compute Nodes".

A Partition driver is a software component dealing with transition requests: it performs the needed steps to convert a host from WN to CN or from CN to WN. Note that nodes are not removed from one cluster, nor joined the another; they are simply made active on the target cluster and inactive in the origin cluster.

To do so:

  • At cloud side

    • Compute Nodes are Enabled / Disabled using OpenStack APIs
  • At Batch side

    • Worker Nodes are configured to publish a numeric status value (External Load Index) named dynp, indicating the partition they belong to.

    • LSF Master is configured to alter job parameters at submission time, adding the requirement for a node having the correct value of dynp

Deployment instructions

Assumptions

This guide assumes that:

  • A properly configured Batch System instance is operational (currently LSF, version 7.x or newer)

  • A working OpenStack instance is in place, with a Cloud Controller as privileged member of the LSF cluster.

  • Cluster members (WN and CN) have read access to a shared filesystem ( mounted as

    LSF_TOP=/usr/share/lsf

    for the sake of example )

    In next sections we describe the Batch side and Cloud side deployment instructions seperately.

results matching ""

    No results matching ""