In evaluating a given system, it is interesting to note that almost invariably, real world systems outperform simulated systems. Even when all policies, reservations, workload, and resource distributions are fully captured and emulated. What is it about real world usage that is not emulated via a simulation? The answer is the 'user feedback' loop, the impact of users making decisions to optimize their level of service based on real time information.
A user feedback loop is created any time information is provided to a user which modifies his job submission or job management behavior. As in a market economy, the cumulative effect of many users taking steps to improve their individual scheduling performance results in better job packing, lower queue time, and better system utilization overall. Because this behavior is beneficial to the system at large, system admins and management should encourage this behavior and provide the best possible information to them.
There are two primary types of information which help users make improved decisions. Cluster wide resource availability information and per job resource utilization information.
Maui provides a number of informational commands which help users make improved job management decisions based on real-time cluster wide resource availability information. These commands include showbf, showgrid, and showq. Using these commands, a user can determine what resources are available, and what job configurations statistically receive the best scheduling performance.
A job's resource requirement specification tells the scheduler what type of compute nodes are required to run the job. These requirements may state that a certain amount of memory is required per node or that a node have a minimum processor speed. At many sites, users will determine the resource requirements needed to run an initial job. Then, for the next several years, they will use the same basic batch command file to run all of their remaining jobs even though the resource requirements of their subsequent jobs may be very different from their initial run. Users often do not update their batch command files even though these constraints may be unnecessarily limiting the resources available to their jobs for two reasons: 1) users do not know how much their performance will improve if better information were provided. 2) users do not no exactly what resources their jobs are utilizing and are afraid to lower their job's resource requirements since doing so might cause their job to fail.
To help with determining accurate per job resource utilization information, Maui provides the FEEDBACKPROGRAM facility. This tool allows sites to send detailed resource utilization information back to users via email, to store it in a centralized database for report preparation, or use it in other ways to help users refine their batch jobs.