Basic rules

  • always check the location where you read/write files and space availability. See df command.
  • AVOID using nfs file systems/mounts for your simulations, especially when running on multiple nodes.
    This will double sometimes your execution time since for every read/write operation an extra IOwait time is needed to transfer information over the network.
    Use a location mounted on a local disk, usually /export/work/export-b/work/ or /scratch (on HPC nodes) and move your data after finishing in a desired location.

To learn

Commands

Local Linux Cluster

  • info about the cluster status using ganglia (works only internally).  If you choose then "cluster linux" , you see in "choose a node" the available computers.
  • from outside, you have first to connect via ssh  to login.stud.ntnu.no (or login.ansatt.ntnu.no). From there you can connect in the next step to a local computer.
  • use nice/renice to lower the priority of your processes
  • you can ran processes in parallel using MPI. Multithread libraries in C++, such as pthread, are also available.


Ganglia

Ganglia provides a web-based user interface that can be used to view activity and available processes on the Linux cluster. It can only be accessed internally. To do this, use ssh to access one of the computers in the cluster.

If one wants to use a graphical web browser, such as FireFox, one needs to use the optional argument -X (enables X11 forwarding). The arguments -Y (enables trusted X11 forwarding) and C (Requests compression of all data) can speed up the web browser.

Alternatively, you can use text-based web browsers such as "links" or "lynx".

Example:

ssh -XYC username@login.stud.ntnu.no
ssh -XYC computername.phys.ntnu.no
# Open browser
firefox -no-remote -no-xshm http://bird/ganglia/
# links http://bird/ganglia/
# lynx http://bird/ganglia/


Software


  • No labels
Write a comment...