Request to reintroduce or provide a mechanism for CPU affinity control for GPU devices. With the removal of the affinity config option, there's no way to specify which CPU cores the GPU should use, leading to PCORE oversubscription and underutilization of ECORES in certain configurations.
### Request Description With the removal of affinity config option currently there is no way to tell GPU device which CPU cores it should utilize and it always preffers PCORES. This leads to a situation of PCORE oversubscription in devices with low PCORE count (think of 2+8 configuration). So all PCORES end up 100% utilized, ECORES are not used at all. I could not find a way to fix it, setting affinity via syscall manually does not work, SCHEDULING_CORE_TYPE config option is not supported on a GPU device. Why is it important? When running multiple workloads which use GPU, but still need some CPU work achieved performance is lower than it could be possible. In my test of whisper speech to text benchmark, current implementation is leaving something like 10%-15% of performance on the table, because: - PCORES are overutilized - ECORES are not used at all - GPU is underutilized, because it waits for PCORES ### Feature Use Case Implementation of this feature would position us better ag