Release Notes for HPC Pack 2016 Update 3

  • Article

These release notes address late-breaking issues and information for the high-performance computing (HPC) cluster administrator about Microsoft HPC Pack 2016 Update 3.

Download and install Microsoft HPC Pack 2016 Update 3

HPC Pack 2016 Update 3 is available for download from the Microsoft Download Center. Download it to your local machine which will act as the local head node. After the download, right-click the zip file of the installation package and click Properties to view the file properties. Click Unblock if there is a security warning that indicates the file might be blocked. Then, extract the installation package files to a local folder and run Setup.exe.

Fixed the following issues

Performance and reliability improvements

  • Fix HPC services may not auto restart issue.
  • Fix port exhaustion in Web Portal (w3wp) caused by HPC diagnostics service.
  • Fix scheduler TCP connection leak.
  • Fix management service connection and port leak on 6730.
  • Fix scheduler crash due to unhandled exception when deserializing network stream.
  • Fix HpcBrokerWorker.exe crash for ArgumentNullException.

Management

  • Fix the issue that compute/broker node is taken offline frequently when running on a Hyper-V virtual machine with dynamic memory feature enabled.
  • Fix management service WCF connection error in Kerberos only environment.
  • Fix the issue that sometimes "Version" and "OS version" of HPC nodes are not correct.
  • Fix node provisioning failure for 1000+ nodes by increasing the max concurrent numbers in service throttling.
  • Fix task environments can not be removed issue in HPC Cluster Manager and HPC Job Manager.
  • Fix HPC Cluster Manager or HPC Job Manager crash due to null reference issue.

SOA runtime

  • Fix object disposed fault message triggered by address already in use error at service host. Scheduler
  • Fix the issue that a queued job with required nodes would take more idle nodes besides the reserved which could block other queued jobs to run.
  • Fix Azure burst time out and nodes end with zombie state.

Linux node manager

  • Fix a bug that 'Cores In Use' is not correct.
  • Fix a bug that HPC Pack environment variables will be lost if CCP_SWITCH_USER is set on CentOS.
  • Fix a bug that memory is limited to the first NUMA node when using cgroup.
  • Fix a bug that zombie task cleanup would fail when node manager starts.
  • Fix a bug that processes in task are not actually terminated after task canceling when cgroup is not enabled.
  • Fix node manager crash issue when monitoring GPU metrics.
  • Change the owner of home directory, which is created by Linux nodemanager, to the user instead of leaving it root.
  • Fix mutual trust failure issue when user home directory is not set as default path.