gintan/intro-to-hpc

Intro to HPC lesson materials

intro-to-hpc's Issues

Feedback on HPC Intro and HPC Carpentry

Greetings from the HPC Carpentry Steering Committee!

We are reaching out to you because this repository is a fork of the HPC Intro lesson. The community that maintains your "upstream" repository is currently involved in an effort to become an official lesson program of The Carpentries(tm). We are reaching out to various forks in an effort to identify and build a larger and more integrated HPC Carpentry community.

Part of this process is to show that the upstream lesson is being used, and that there is a pool of potential instructors interested in giving the lesson. This is a big part of showing that there is the scope to grow a wider community around HPC Carpentry.

Is someone from your team is interested in collaborating in this larger community? If so, please indicate yes in this issue. Forks like yours can potentially add value to the broader HPC Carpentry community by expanding the base of instructors, providing feedback on your experience in teaching the material, and providing edits to the material through pull requests or issues, as well as participation in community governance.

Parallel concepts - impression from the first course

I felt (no feedback evidence of this), that the class was struggling to move from the cooking analogy to the job submission concepts. The cooking analogy works and should be kept. The only way I have come up with of getting across the concepts is something of a history lesson, combined with examples.

1. Old days computers had 1 core, and lots of users. Weren't really parallel but looked like it via time slicing. The OS started and stopped processes very quickly.
2. As more CPUs were included in one machine the process became the first vehicle for parallel software. Two processes would work on different parts of the same problem and OS ensure that they ran on different CPUs at the same time. Processes can communicate.
3. Also, people began building clusters - collections of machines on a fast network, possibly with one CPU each. Layer of software introduced so that processes running on different machines can communicate in order to work on the same problem. One of the commonly used software layers is called MPI.
4. As multicore machines became more common, threads were introduced. Threads live inside a process and share memory. This makes communication more efficient than MPI or interprocess communication. The OS knows about threads and ensures that threads run on different cores when possible.

Applications need to be written especially to support different types of parallelism, and you need to know which is being used in order to submit properly with SLURM.

Then onto how tasks, cpus-per-task and nodes map to these concepts.

Difficulties submitting massive MPI jobs, etc