Hello,
I have a system with MALI G78 AE GPU which has 4 slices in 4 different partitions, each slice with 2 shader cores.
I would like to know if it is possible to assign a time slice for a task to have access to the GPU partition in the system, and if so, how this can be done?
Example: Task A and Task B execute in the same partition on the system, am I capable to assign 10ms to Task A and then 20ms to Task B and so on until the tasks are finished?
How the context switch is driven in the MALI GPU? Is it something configurable in its arbiter?
Thanks, Luca.
Hi Luca,
The Arbiter provides a simple round-robin fair scheduling policy based on the mali_arbiter.ko's request_timeout parameter.
The stack is provided as a reference, so it is expected that both the configuration (module parameter and sysfs interface) and scheduling policy is replaced with a use-case and system specific implementation.
Hi Richard,
Thanks for your reply on this case.
I don't have any further questions, but if anyone is interested in modifying their default request_timeout value for the context switch on the GPU, this can be done dynamically through the following command:
$ echo <new_value> | sudo tee sys/module/mali_arbiter/parameters/request_timeout
<new value> should be in ms and sudo may not be necessary. Example: echo 10 | tee (...)
Kind regards, Luca.