Computing

This is an old revision of the document!

---

Cluster CPU/Memory Usage Report

This table shows cluster cpu/memory usage for qsub jobs in terms of “Slot Equivalent Time” by periods of 'day' and 'division'. qlogin jobs/sessions are not reported.

The goal of this report is to help with deciding what how many slots per “High-speed Slot Group” to request for your group. For a discussion on slots and Slot Groups, click here.

NOTE Slot Group quotas have not yet been implemented on the cluster, but will be very soon. Once this is done, each job will be assigned to an SGE project (with the same name as a slot group) to track and limit by Slot Group quotas. For users who work with multiple groups or labs, this means their quotas will then be set based on the project they assign to their jobs, allowing them to work flexibly with multiple labs. However, the report below is from jobs that do not include that information, so users who work within multiple groups have their usage below reported only as a part of their 'home' group/lab.

# of non-zero days This column shows how many days of the reporting period had at least one job running. All statistics are computed only over days or divisions (see below) during which one of more jobs were run. So for example the reported average number of jobs per day is computed only over “non-zero” days. If you ran 100 jobs only on one day within the reporting period, your daily job average will be 100.

Slot Equivalent A “Slot Equivalent” (SE) represents a job's fractional SGE quota usage in units of either one slot or 6GB of RAM (whichever is greater for a job). We use 6GB because for each slot in a user's quota, 6GB of RAM quota is allotted.

Examples:

Slot Equivalent Time The “Slot Equivalent Time” (SET) is the period of time a job runs, multiplied by the SE of the job, reported in units of days or 'divisions'.

SET = SE * job-duration

So a job that runs for 8 hours with 1 SE is reported as running for 0.33 SET by day. A job running for 16 hours with 2 SE is reported using 1.33 SET by Day (2 SE * 16 hours / 24 hours-per-day) ). SET is also reported by 'division', a period shorter than a full day (the value of which is listed below). For a division period of 4 hours then, a 6 hour job with 1 SE would be reported as 1.5 SET by division ( 1 SE * 6 hours / 4 hours-per-division).

If you ran 1 job that lasted 24 hours, you'd get an SET by Day of 1, and an SET by Division of 1. But if you ran 24 1-hour jobs that all ran within the same 4-hour window, you'd still get an SET by Day of 1, however the SET by Division would be 24. This would show your jobs were more concentrated in a smaller window of time.

The purpose of this metric is to get a useful idea of how many SE units are used at once by a group over useful periods of time, and so have a good idea of what kind of slot-group quota the group needs.

Long jobs Jobs that last longer than a day or a division are spread out over as many days or divisions they straddle.

How to use this report The most important values are the first three columns: '# of non-zero days', 'Slot Equiv. Time By Day' average and standard deviation. Consider how many days your group had jobs running, and then the SET values. For example if your number of non-zero days is high, you may want to add the average SET by Day to its standard deviation to allow users to generally run the same number of jobs as they have in the past. If your number of non-zero days is low, consider just the SET by Day average to save on Slot Group fees. You can also consider the 'Slot Equiv. Time By Div' data. If these values are significantly higher, it probably indicates users are running shorter jobs during the same time periods, e.g. during the work day. In that case you may want to use these numbers to decide how many slots to request.

== REPORT == (note, I will refine the output formatting in future reports)

Period Begin: Sat Jun 4 17:35:37 EDT 2016
Period End : Fri Nov 11 16:35:37 EST 2016
Days: 160
Hours per Division: 4 (for stats reported under “Slot Equiv. Time By Div.”)