Best Practices for High Throughput Jobs

High throughput jobs are high volume computing jobs whose typical computing duration is similar to or even shorter than the time required to receive, schedule, start, and stop them. In addition to the general considerations for high volume jobs, high throughput jobs can strain the capabilities of Argon's home directory servers. In some cases, this results in failed jobs among all high throughput jobs on the system, and slow performance for other users.

Description of the problem

Home directories are stored on separate file servers and connected via network to login and compute nodes using NFS. This design works well for most cluster usage, but when many compute nodes simultaneously try to write, read, or delete files in home directories, access becomes slow for all users and some jobs fail. The job failure errors typically state either that a file or directory does not exist or that access to a file or directory was denied, even though the file/directory does exist and the user has the appropriate permissions to access it. For example:

Some high throughput workloads can generate enough traffic to significantly increase likelihood of this occurrence depending on other user job activity, or even cause this condition by themselves. Also, by hypothesis any "high throughput" workload has a relatively higher incidence of its own jobs failing when the condition occurs.

Most common case

The default behavior is for the job scheduler (Sun Grid Engine) to place redirected output files in the user's home directory. (Files ending in .onnnn and .ennnn , where nnnn is the job number). When a user starts many small jobs simultaneously, a large number of compute nodes will attempt to create .o and .e files and log the output and error streams from each job. The many NFS operations this entails create a very high load on the NFS server. This is the most common case, but not the only one which can trigger the problem.

Other cases

Other cases involve reading or writing many files (sometimes thousands) at the start or end of each job, or even leaving many files open inside each job so that they are cleaned up suddenly at the same time when the function (or entire program, or entire job) ends. A high throughput workload containing many such jobs (sometimes thousands) can multiply these effects to problematic proportions very easily.

Mitigation strategy

The problem can be mitigated by reducing the number of requests sent by compute nodes to the home directory server. We have worked with several people to implement the following mitigation options and achieved a successful outcome.  

1. Redirect standard output and standard error streams as follows:

   1. If you don't want to examine the streams, you can discard them both: join error into output and send output to /dev/null by using -j y -o /dev/null. (If you have the -o option already defined in your job script, you will need to use this instead.)

   2. If you do want to examine one or both streams, first redirect the corresponding files into /localscratch while the job is running. Then at the end of the job, move the file(s) to network storage accessible outside your job script for later review or processing.

      To keep standard error, redirect the .e file to /localscratch using -e /localscratch. Otherwise join it into the .o file using -j y or discard it using -e /dev/null.
      To keep standard output, redirect the .o file to /localscratch using -o /localscratch. Otherwise discard it using -o /dev/null.

      At the end of the job, move whatever files you didn't discard off the compute node's /localscratch drive into a location you can access later by adding the line below to the end of your script. In this example, I am collecting the .o and .e files into a subdirectory under my home directory called "output", but you can move them wherever you prefer:

      for non-merged output

      mv $SGE_STDOUT_PATH ~/output
      mv $SGE_STDERR_PATH ~/output
      

      Note: If you merge the error stream into the output stream (that is, you specified the -j y option), there will be no .e file, so you only need to move the .o at the end of the job.

2. Do not load environment modules with each job or array task. Take advantage of the environment passing to jobs that is done by default. See Environment modules for HTC/HVC jobs for more information.
3. If your high throughput jobs need to read or write a large number of files, you should add commands to your job script to first copy any input files to /localscratch, accordingly modify the location where each processing step or program in the job expects input files and writes its intermediate and final result files, and add commands at the end of the job script to clean up and collect result files onto network storage (typically your home directory). The details depend on what your job needs to do, which programs it uses, and generally how it operates. If you would like further advice or assistance, please email us at research-computing@uiowa.edu.

Why this works

The /localscratch area is space that is local to each node. Therefore, the load of constantly accessing and changing these files is distributed onto the compute nodes running your jobs instead of being focused on a single network server.  

If you have any questions

Please let us know if you have any questions or if you have suggestions on how we can improve this document. Send an email to research-computing@uiowa.edu