Cray User Group Directory of User-submitted Software

Title	Job Mix Daemon or Dynamic Job Mix Scheduler
Revision	July, 1991
Purpose	Dynamically sets optimal NQS job mix timeslices long jobs
Package type	program
Application area	Operating System
Abstract	Jobmixd is a user level daemon that is used to control NQS jobs. Based on the current job load, it determines the optimal jobmix to run. Also, in order to prevent long running jobs from blocking the flow of jobs in queues, it time-slices running jobs via the suspend and resume mechanism.
Submittor	Wayne Schroeder
Contact	Wayne Schroeder
Phone	(619) 534-5065
Email	schroeder@sdsc.edu
Company/ Institution	San Diego Supercomputer Center
CUG Site Code	SDSC
Distribution method	contact the following person: schroeder@sdsc.edu
URL
Distribution restrictions	UNICOS sites only
Copyright	licensed, not-for-resale, was developed as a joint SDSC/CRI project.
License required	yes
License fee	$200
Hardware	Cray XMP, YMP
Operating System	UNICOS 5.1
Compiler(s)	pcc
Program support	On a continuous basis. For an additional $4,000 fee, we will provide consulting to assist a site with an installation.
Documentation	README file, Makefile, man pages, programmer documentation
Portability to other UNIX systems	Not applicable and not allowed by license (SDSC retains rights).
Additional comments	Requires a NQS queue complex configuration as follows: All NQS queues controlled by the scheduler are grouped into a few complexes based on job sizes (i.e. queues belonging to the same complex controlled by the scheduler must have the same size). Job control derives from adjusting the complex run_limit and user_limit. Input in the jobmixd.cf file defines which complexes are controlled by the scheduler. The total number of complexes controlled by jobmixd must be less than or equal to 10. For complexes controlled by the scheduler, there should be no overlapping of membership. i.e. a queue cannot belong to more than one complex controlled by the scheduler. It is permissible to have two or more complexes controlled by the scheduler to have the same size. e.g. on our system, complexes "c-4" and "standby" both have a size limit of 4 mw. It is permissible to define complexes that are not controlled by the scheduler. Also, the rule of no overlapping member does not apply for complexes that are not controlled by the scheduler. e.g., we defined a complex (not controlled by jobmixd) named "c-disk" for jobs that use large amount of disk space. Members of the "c-disk" complex are also members of complexes that are controlled by the scheduler. The "c-disk" complex was create so that we can control the number of large disk jobs running in the system.