1. Do not construct jobs on the big cluster that do too much NFS I/O. Generally, you should set things up such that only the output goes to the file servers, and craft your programs so that the output is dense. Use the local disk /scratch and the semi-local disk /iscratch/i for input. Put temporary files in /tmp.
2. Avoid running long compute-intensive jobs on hgwdev whenever possible. We all compile and test the browser on this machine.
3. Try to set up jobs that run on the cluster to take between 5 seconds and an hour on average. Shorter jobs will waste cluster bandwidth for technical reasons I won't go into now. Longer jobs make sharing the cluster more awkward. The cluster shares between users by awarding the next free CPU to the user with the least jobs running. It does not preempt jobs. If you have 100 or less jobs of longer duration, it is OK to run them though.
4. Try to do I/O intensive jobs, such as sorting cluster output, on the file server itself. This is usually the machine eieio or kkstore. To determine this, do a df in the directory in which you are interested.
5. Be sure to check that you have enough free space on disk before launching jobs that will use a lot of free space.
6. Avoid making more than 20,000 files in a single directory. The system, particularly under NFS, gets slow on huge directories.
7. If you generate a track, you are responsible for updating the corresponding make*.doc file to describe the steps you took to create it.