2.7 Priorities and Preemption

Condor has two independent priority controls: job priorities and user priorities.

2.7.1 Job Priority

Job priorities allow the assignment of a priority level to each submitted Condor job in order to control the order of their execution. To set a job priority, use the condor_prio command; see the example in section 2.6.4, or the command reference page on page . Job priorities do not impact user priorities in any fashion. A job priority can be any integer, and higher values are better.

2.7.2 User priority

Machines are allocated to users based upon a user's priority. A lower numerical value for user priority means higher priority, so a user with priority 5 will get more resources than a user with priority 50. User priorities in Condor can be examined with the condor_userprio command (see page ). Condor administrators can set and change individual user priorities with the same utility.

Condor continuously calculates the share of available machines that each user should be allocated. This share is inversely related to the ratio between user priorities. For example, a user with a priority of 10 will get twice as many machines as a user with a priority of 20. The priority of each individual user changes according to the number of resources the individual is using. Each user starts out with the best possible priority: 0.5. If the number of machines a user currently has is greater than the user priority, the user priority will worsen by numerically increasing over time. If the number of machines is less then the priority, the priority will improve by numerically decreasing over time. The long-term result is fair-share access across all users. The speed at which Condor adjusts the priorities is controlled with the configuration variable PRIORITY_HALFLIFE , an exponential half-life value. The default is one day. If a user that has user priority of 100 and is utilizing 100 machines removes all his/her jobs, one day later that user's priority will be 50, and two days later the priority will be 25.

Condor enforces that each user gets his/her fair share of machines according to user priority both when allocating machines which become available and by priority preemption of currently allocated machines. For instance, if a low priority user is utilizing all available machines and suddenly a higher priority user submits jobs, Condor will immediately take a checkpoint and vacate jobs belonging to the lower priority user. This will free up machines that Condor will then give over to the higher priority user. Condor will not starve the lower priority user; it will preempt only enough jobs so that the higher priority user's fair share can be realized (based upon the ratio between user priorities). To prevent thrashing of the system due to priority preemption, the Condor site administrator can define a PREEMPTION_REQUIREMENTS expression in Condor's configuration. The default expression that ships with Condor is configured to only preempt lower priority jobs that have run for at least one hour. So in the previous example, in the worse case it could take up to a maximum of one hour until the higher priority user receives a fair share of machines. For a general discussion of limiting preemption, please see section 3.5.9 of the Administrator's manual.

User priorities are keyed on <username>@<domain>, for example johndoe@cs.wisc.edu. The domain name to use, if any, is configured by the Condor site administrator. Thus, user priority and therefore resource allocation is not impacted by which machine the user submits from or even if the user submits jobs from multiple machines.

An extra feature is the ability to submit a job as a nice job (see page ). Nice jobs artificially boost the user priority by one million just for the nice job. This effectively means that nice jobs will only run on machines that no other Condor job (that is, non-niced job) wants. In a similar fashion, a Condor administrator could set the user priority of any specific Condor user very high. If done, for example, with a guest account, the guest could only use cycles not wanted by other users of the system.

2.7.3 Details About How Condor Jobs Vacate Machines

When Condor needs a job to vacate a machine for whatever reason, it sends the job an asynchronous signal specified in the KillSig attribute of the job's ClassAd. The value of this attribute can be specified by the user at submit time by placing the kill_sig option in the Condor submit description file.

If a program wanted to do some special work when required to vacate a machine, the program may set up a signal handler to use a trappable signal as an indication to clean up. When submitting this job, this clean up signal is specified to be used with kill_sig. Note that the clean up work needs to be quick. If the job takes too long to go away, Condor follows up with a SIGKILL signal which immediately terminates the process.

A job that is linked using condor_compile and is subsequently submitted into the standard universe, will checkpoint and exit upon receipt of a SIGTSTP signal. Thus, SIGTSTP is the default value for KillSig when submitting to the standard universe. The user's code may still checkpoint itself at any time by calling one of the following functions exported by the Condor libraries:

ckpt()()

Performs a checkpoint and then returns.

ckpt_and_exit()()

Checkpoints and exits; Condor will then restart the process again later, potentially on a different machine.

For jobs submitted into the vanilla universe, the default value for KillSig is SIGTERM, the usual method to nicely terminate a Unix program.