SUBMITTING JOBS TO ARTEMIS FROM MATLAB

Transcription

1 INFORMATION AND COMMUNICATION TECHNOLOGY SUBMITTING JOBS TO ARTEMIS FROM MATLAB STEPHEN KOLMANN, INFORMATION AND COMMUNICATION TECHNOLOGY AND SYDNEY INFORMATICS HUB 8 August 2017

2 Table of Contents GETTING STARTED... 2 GET AN ARTEMIS ACCOUNT... 2 INSTALL MATLAB R2017A... 2 IMPORT THE ARTEMIS CLUSTER PROFILE... 2 ARTEMIS CLUSTER PROFILE CONFIGURATION... 2 VALIDATE THE ARTEMIS CLUSTER PROFILE... 4 VALIDATION FAILURE BUG FIX FOR LINUX USERS... 6 SUBMITTING WORK TO ARTEMIS VIA MDCS... 7 SUBMITING WORK TO ARTEMIS USING THE BATCH COMMAND... 7 ALTERNATIVE WAYS OF SUBMITTING WORK TO MDCS... 8 MONITORING JOBS SUBMITTED TO ARTEMIS VIA MDCS... 9 GETTING RESULTS FROM FINISHED JOBS DOWNLOADING DATA FROM ARTEMIS TO YOUR LOCAL COMPUTER OR RDS Stephen Kolmann 8 August 2017 Page 1

3 Getting Started We have installed Matlab Distributed Computing Server (MDCS) on the University s High- Performance Computing cluster Artemis. This guide will teach you how to configure your Matlab R2017a client to submit jobs to Artemis, and it will show basic examples of submitting and monitoring submitted jobs. Get an Artemis Account You will need to have Artemis access before you can submit jobs to Artemis using your Matlab client. Install Matlab R2017a Only Matlab R2017a can submit Matlab jobs to Artemis, so you will need to install it on your local computer. The University of Sydney has a site license for Matlab and provides access to this license at no cost to Sydney University researchers. You can download and install Matlab R2017a from the softserv website. Import the Artemis cluster profile The Artemis cluster profile is a set of Matlab scripts provided by us that automates submitting your Matlab scripts to Artemis. Follow these instructions to import the Artemis cluster profile into your Matlab client: Download artemis-mdcs.zip from the softserv website and extract the contents. Open Matlab R2017a on your computer. Click the Home tab, select Parallel, then select Manage Cluster Profiles In the Cluster Profile Manager, click the Import button and select the Artemis.settings file. If you successfully imported the Artemis cluster profile, you should see a new cluster profile called Artemis in the Cluster Profile manager. Artemis cluster profile configuration The imported Artemis cluster profile needs additional configuration before use. Follow the steps below to complete the configuration: Select the imported Artemis cluster profile in the Cluster profile manager window. Click the Edit button, either in the lower-right hand corner, or in the menu bar at the top of the window. In the Scheduler Integration section, set the Folder containing scheduler integration scripts option to the path to the artemis-integration-scripts folder you extracted earlier. In the additional properties for integration scripts section: Set Project to your abbreviated project name as specified in your RDMP. Set RemoteJobStorageLocation to a directory on Artemis where you would like your job s files to be stored. For information about Artemis storage, see the Data storage on Artemis section of the Artemis user guide. Adjust the Walltime option to a value appropriate for your job. Walltime is the maximum amount of time your job will run on Artemis. If your job exceeds this walltime, it will be terminated and you will lose all unsaved progress. It is best to overestimate the walltime if you are unsure of how much you need. Stephen Kolmann 8 August 2017 Page 2

4 Other valid options are: Mem: the amount of RAM (memory) per core (or CPU) your job needs. It is best to overestimate if you don t know how much you require. Ngpus: the number of GPUs required. The maximum you can request is two. Only use this option if you know your job can use GPUs. Queue: the Artemis job queue your job will be placed in. If you are unsure of what these are, leave it blank. For information about Artemis queues, see the Artemis user guide. Important note: MDCS jobs cannot be submitted to the scavenger queue. Jobs submitted to scavenger will be rejected with this error qsub: Job violates queue and/or server resource limits. AdditionalSubmitArgs: Additional PBS options for your submitted job. In the Files and Folders section, in the AttachedFiles option, you may specify a directory or multiple directories on your local computer that will be copied to Artemis before your job starts. You may leave this blank if not required. In the AdditionalPaths box in the Files and Folders section, you may specify a folder or multiple folders on Artemis that contains input files needed for your job that you copied to Artemis in advance. You may leave this blank if not required. When you have finished editing the Artemis cluster profile, click Done in the bottom right-hand corner of the window. A sample Artemis cluster profile is shown on the next page. Stephen Kolmann 8 August 2017 Page 3

5 Validate the Artemis cluster profile To ensure the Artemis cluster profile is working, run the Cluster Profile Validation: Select the Artemis Cluster Profile in the Cluster Profile manager, then Click the Validate icon in the menu bar at the top of the Cluster Profile Manager. The validation window will appear. Before the validation tests can run, you must Authenticate to the MDCS server. First, enter your UniKey in the Enter the username for mdcs.hpc.sydney.edu.au box, then click OK. Stephen Kolmann 8 August 2017 Page 4

6 Next, when it asks if you wish to use an identity file to login to mdcs.hpc.sydney.edu.au, answer No. Finally, enter your password in the next dialog box, then click OK. Stephen Kolmann 8 August 2017 Page 5

7 If the test was successful, all tests should pass except the Parallel pool test (parpool). Validation failure bug fix for Linux users Matlab R2017a MDCS validation on Linux fails with a rather cryptic error about timezones and a script called datetime.m. If you encounter this bug and wish to fix it, follow the following steps: 1. Back up your datetime.m file. The path to datetime.m is: MATLABROOT/toolbox/matlab/timefun/@datetime/datetime.m, where MATLABROOT is the directory where Matlab R2017a is installed on your computer. 2. Download the patched datetime.m file, and rename it to datetime.m. 3. Copy the patched datetime.m file to MATLABROOT/toolbox/matlab/timefun/@datetime/datetime.m. Stephen Kolmann 8 August 2017 Page 6

8 Submitting work to Artemis via MDCS You can submit non-interactive (batch) work to Artemis via MDCS. The following examples will use an example Matlab script called quadrature.m: n= ; a=0.5; b=1.0; q=0.0; w=(b-a)/n; parfor i=1:n q = q + w*bessely(4.5,((n-1)*a+(i-1)*b)/(n-1)); end q After saving this script to your local computer, you can run it with the following command in the Matlab Command Window: >> quadrature Submiting work to Artemis using the batch command The quadrature.m script can be run on Artemis using the batch command in the Matlab Command Window: >> job = batch('quadrature','profile','artemis') Stephen Kolmann 8 August 2017 Page 7

9 Using multiple cores The quadrature.m script can be run using multiple cores since it has been explicitly parallelised with a parfor loop, you can request multiple cores using the Pool option to batch: >> job = batch('quadrature','profile','artemis','pool',3) The above command initialises a parallel computing pool with 3 cores that do the computation, and one additional head core that executes the script. Therefore, the above example requests 4 cores. Important note: You can have up to 12 cores worth of work running simultaneously. Any work that would put you over this 12 core limit will remain queued until there are enough cores free to run your job. Alternative ways of submitting work to MDCS There are other possible ways of submitting work to Artemis via MDCS. For full details about other methods, refer to Mathwork s Parallel Computing Toolbox Documentation. Stephen Kolmann 8 August 2017 Page 8

10 Monitoring Jobs submitted to Artemis via MDCS Jobs submitted to Artemis via MDCS can be monitored in using the Job Monitor in Matlab R2017a on your own computer. To open the Job Monitor, click the Home tab, click the Parallel button, then click Monitor Jobs. The Job Monitor window should open inside Matlab. Next, select the Artemis profile in the Select profile tab: You will need to enter your Unikey credentials again: Enter your UniKey into the first pop-up box. Select No to the second pop-up box asking for an identity file. If a third pop-up box appears, enter your password. If you successfully authenticated, you will see a list of jobs that are either queued, running or finished. Stephen Kolmann 8 August 2017 Page 9

11 To interact with the jobs, you can right-click them and select one of: Cancel: terminates the job or removes it from the queue. Delete: terminates the job, removes it from the queue and deletes it from the Job Monitor. Any data associated with the job will still remain on Artemis. Show details/errors/warnings: Shows details/errors/warnings for the job. Fetch outputs or load variables: Downloads all data stored in Matlab variables for that job and loads them into your local workspace. Important Note: Due to the nature of Artemis s job scheduling cycle, it may take up to 20 minutes (worst-case scenario) to perform cancel job or delete job operations. Getting results from finished jobs After your job finishes, you can download the final variables from your job by either rightclicking your job in the Matlab Job Monitor window and selecting Fetch outputs or Load variables (whichever one applies), or alternatively, in the Matlab Command window type: mycluster = parcluster('artemis'); job = mycluster.findjob('id',20) load(job) This would load the variables of a job with ID 20 into your Matlab client. Downloading data from Artemis to your local computer or RDS If, in your cluster profile, you set your RemoteJobStorageLocation to /project/pandora, then your job data would be stored in /project/pandora/abcd1234/job<jobid>, where PANDORA is the abbreviated project name, abcd1234 is your UniKey and <JobID> is the Job ID assigned to your job by Matlab. You can download this data from Artemis to your local computer or from Artemis to RDS using a Secure File Transfer Protocol program such as WinSCP, FileZilla or Cyberduck. For help using these programs, see the Artemis User Guide. Stephen Kolmann 8 August 2017 Page 10