Page faulting and selective preemption

A method of workload management on a parallel processor, the method comprising:

• in response to receiving a notice that a number of page faults associated with a first work group on a first compute block within the parallel processor has exceeded a threshold, placing the first compute block into a preemption state;
• determining if a first pending workgroup is available for execution on the first compute block;
• preempting the first work group on the first compute block when the first pending workgroup is available for execution on the first compute block;
• and
• executing a second work group on a second compute block within the parallel processor while preempting the first work group, wherein the first compute block and the second compute block each include multiple processing clusters.

A data processing system comprising:

• a memory device;
• and
• a parallel processor coupled to the memory device, the parallel processor comprising a processing array including multiple compute blocks, each compute block including multiple processing clusters to perform parallel operations, the parallel processor additionally including a scheduler coupled with the processing array, wherein the scheduler is to:
• place a first compute block of the multiple compute blocks into a preemption state in response to receipt of a notice that a number of page faults associated with a first work group on the first compute block has exceeded a threshold;
• determine if a first pending workgroup is available for execution on the first compute block;
• preempt the first work group on the first compute block when the first pending workgroup is available for execution on the first compute block;
• and
• execute a second work group on a second compute block of the multiple compute blocks during preemption of the first work group.