Where to Run the PCT
There are two platforms on which to run the PCT: locally on a PC or remotely on a cluster. Which is more appropriate for you depends on accessibility and job requirements.
- Run locally on a multicore desktop or laptop.
If you have access to a multicore PC or laptop, this is suitable for code development or for applications that can take advantage of up to 12 cores with MATLAB 2011b (or 8 cores with older versions).
- If you don’t have MATLAB installed on your machine, please check if you qualify to install it on your local machine.
- The PCT works on single and multicore computers. However, speed up depends on the number of physical cores.
- A worker is MATLAB’s term for a processor.
- If you are using MATLAB licensed to BU, your computer needs to either be physically on campus or, if accessing from off-campus (e.g., home), log in via VPN. Running MATLAB will trigger an automatic process to check out the appropriate MATLAB and PCT licenses. There is no action required by you.
- Run remotely on SCV’s Katana Cluster.
This is the only way to run PCT jobs with more than 16 processors, on multiple nodes. With the appropriate batch submission procedure, you may also use this server to run jobs as “local” on a single node. At present, the biggest nodes on Katana each has 12 cores and hence that is the most you can use for a “local” job.
- To access Katana requires SCV userid. If you don’t already have one, please check for eligibility.
- SCV maintains four clusters. Only the Katana Cluster supports the MATLAB PCT.
- Before running MATLAB parallel tasks on Katana’s remote nodes for the first time, you must set up a MATLAB configuration file.
- A MATLAB PCT user may use up to 32 workers in a single job or a total of up to 32 workers for multiple concurrent jobs.
- The Katana login node has 4 processors. You can develop codes and run light, short-duration, parallel jobs on it.
Caveat: timing collection using tic/toc or other timing devices may be unreliable on the login node because its four processors are shared among all logged in users.