Jobs can run within a session window. The scheduler prioritizes the jobs belonging to that session. Qiskit Runtime primitives transparently take advantage of other session features like shared caching used by the jobs. This helps primitives run as efficiently as possible in the quantum data center and helps users experience a faster turnaround on results for their workload.
How to run a job in a session¶
In Qiskit, you can associate a primitive with a session by using a context manager.
When you call a primitive by using a context manager and that job runs, a session is started (or the job is run in an active session if a session is already active). For example, the following code uses a context manager to call the Estimator primitive:
When you start a runtime job with the jobs API, by default it does not run in a session. You can specify that it should run in a new session (start_session=true) or you can assign the job to run in a current session (session_id=ID of the session that the job is part of). For information about using the jobs API, see the qiskit-ibm-runtime API Reference
Do not start a session inside of a reservation. If you use a session inside a reservation and all the session jobs don’t finish during the reservation window, the pending jobs outside of the window might fail.
How session jobs fit into the job queue¶
For each backend, jobs that are part of a dedicated reservation take priority. If there are no jobs from a provider that has a reservation, jobs that are part of an active session are run. If there are no jobs in an active session, the next job from the regular fair-share queue is run.
A job from the fair-share queue could activate or reactivate a session.
How long is a session active?¶
When a session is started, it is assigned a maximum session timeout value. This timeout value is the same as the initial job’s maximum execution time and is the smaller of 1) the system limit (8 hours for physical systems) and 2) the max_execution_time defined by the program. After this time limit is reached, the session closes permanently and any queued jobs are put into an error state.
Additionally, there is an interactive timeout value. If there are no session jobs queued within that window, the session is temporarily de-activated and normal fair-share job selection resumes. If the scheduler gets a job from the session again, the session is re-activated until its maximum timeout value is reached.