9. Unit Scheduler¶
The class:radical.pilot.UnitManager dispatches compute units to available pilots for execution. It does so according to some schedulin algorithm, which can be selected when instantiating the manager. Momentarily we support two schduling algorithms: ‘Round-Robin’ and ‘Backfilling’. New schedulers can be added to radical.pilot – please contact us on the mailing list `firstname.lastname@example.org‘ for details on support.
Note that radical.pilot enacts a second scheduling step: once a pilot agent takes ownership of units which the unit manager scheduler assigned to it, the agent scheduler will place the units on the set of resources (cores) that agent is managing. The agent scheduler can be configured via agent and resource configuration files (see List of Pre-Configured Resources).
9.2. Round-Robin Scheduler (
The Round-Robin scheduler will fairly distributed arriving compute units over the set of known pilots, independent of unit state, expected workload, pilot state or pilot lifetime. As such, it is a fairly simplistic, but also a very fast scheduler, which does not impose any additional communication roundtrips between the unit manager and pilot agents.
9.3. Backfilling Scheduler (
The backfilling scheduler does a better job at actual load balancing, but at the cost of additional communication roundtrips. It depends on the actual application workload if that load balancing is beneficial or not.
Backfilling is most beneficial for large numbers of pilots and for relatively long running units (where the CU runtime is significantly longer than the communication roundtrip time between unit manager and pilot agent).
In general we thus recomment to not use backfilling
- for a single pilot;
- for large numbers of short-running CUs.
The backfilling scheduler (BF) will only dispatch units to pilot agents once the pilot agent is in ‘RUNNING’ state. The units will thus get executed even if one of the pilots never reaches that state: the load will be distributed between pilots which become ‘ACTIVE’.
The BF will only dispatch as many units to an agent which the agent can, in principle, execute concurrently. No units will be waiting in the agent’s own scheduler queue. The BF will react on unit termination events, and will then backfill (!) the agent with any remaining units. The agent will remain under-utilized during that communication.
In order to minimize agent under-utilization, the user can set the environment variable RADICAL_PILOT_BF_OVERSUBSCRIPTION, which specifies (in percent) with how many units the BF can overload the pilot agent, without waiting for unit termination notices. This mechanism effectively hides the communication latencies, as long as unit runtimes are significantly larger than the communication delays. Thre default oversubscription value is ‘0%’, i.e. no oversubscription.