none
Adding Headnodes to a HPC Cluster RRS feed

  • Question

  • How do I add extra HEADNODES to an existing HPC 2016 cluster. As far as I can see, you specify the headnodes when you create the databases. Is it necessary to recreate the entire HPC 2016 headnode cluster?
    Wednesday, March 13, 2019 9:46 AM

Answers

  • Hi Jacob,

      If it is single Headnode cluster, you're not supposed to add more headnodes into the cluster. This is something we are going to do in HPC Pack 2019 where you can add additional Headnode to a cluster to get High Availability.

      If you're using 3-HN cluster, currently we only support replace existing Headnode when some headnode is down.


    Qiufang Shi

    • Marked as answer by Jacob Hertz Thursday, March 14, 2019 7:34 AM
    Thursday, March 14, 2019 6:24 AM

All replies

  • Hi Jacob,

      If it is single Headnode cluster, you're not supposed to add more headnodes into the cluster. This is something we are going to do in HPC Pack 2019 where you can add additional Headnode to a cluster to get High Availability.

      If you're using 3-HN cluster, currently we only support replace existing Headnode when some headnode is down.


    Qiufang Shi

    • Marked as answer by Jacob Hertz Thursday, March 14, 2019 7:34 AM
    Thursday, March 14, 2019 6:24 AM
  • Sounds like you know what you are talking about - so no way of expanding an existing 4-headnode cluster to a 6-headnode cluster without rebuilding the cluster. It would make more sense to me to have the option/possibility to extend an HA cluster to handle more load - even on headnodes.

    Thanks for the answer

    Jacob Hz


    • Edited by Jacob Hertz Thursday, March 14, 2019 7:50 AM
    Thursday, March 14, 2019 7:34 AM
  • Hi, Jacob,

      For HPC Pack, the major workloads is the scheduler service, and the scheduler service can only run on one instance, thus having more nodes, won't bring any performance benefit for you. And our customer even looking for fewer headnode (2 instead of 3). 

      And if you have thousands of compute nodes, the bottleneck still won't be the headnode, but the backend SQL. If you do have a large cluster to manage (More than one thousand nodes), do contact me as there is some best practice on this.


    Qiufang Shi

    Friday, March 15, 2019 3:20 AM