Session-Based, Estimation-less, and Information-less Runtime Prediction Algorithms for Parallel and Grid Job Scheduling

The default setting of most production parallel job schedulers is FCFS with backfilling. Under this setting, users must supply job runtime estimates, which are known as being highly inaccurate and inferior to system generated predictions. Recent research revealed how to utilize system predictions for backfilling, and this potential performance gain motivates searching for better prediction techniques. We present three prediction techniques using decreasing levels of information as is suitable for the situation at hand. The first is based on "user sessions": continuous temporal periods of per-user work. This algorithm exploits the entire long-term historical data of the workload, along with user runtime estimates. The second is "estimation-less", that is, uses historical data only, relieving users from the annoying need to supply estimates. The third is completely "informationless" and is suitable for cases in which neither historical information nor estimates are available, as happens in some grid environments. We evaluate the algorithms by simulating real data from production systems. We find all of them to be successful in terms of both accuracy and performance.