[Xen-devel] Has anyone ever talked about/implemented scheduling policy for Hyperthread/Multi-core/NUMA?

[Caz Yokoyama <cazyokoyama@xxxxxxxxx>]

Hello Xen developers,

I am new to this mailing list. I examined Credit scheduling and found that it does load balancing between symmetric multi processors (SMP). In the function, csched_load_balance(), it checks whether the physical CPU(PCPU) N+1 has something to do when PCPU N is nothing to do. Note: N is CPU ID. If so, N takes the job of N+1. If N+1 is idling, then N+2 is checked. In this scheme, N+1, N+2, … are checked in this order because csched_load_balance() assumes SMP. On Hyperthread/Multi-core/NUMA, it is better to check the closest neighbor first, i.e. the same Hyperthread CPU first, then the same socket CPU second, then others. To find the closest neighbor first, I am thinking to have a 2 dimensional array. I.e. for CPU N,

Neighbor[N, 0] = CPU ID of the same Hyperthread CPU

Neighbor[N, 1] = CPU ID of the same socket CPU

Neighbor[N, 2] = CPU ID of the same socket CPU

Neighbor[N, 3] = CPU ID of others

Neighbor[N, 4] = CPU ID of others

…….

 

1. Has anyone ever talked about or implemented csched_load_balance() for Hyperthread/Multi-core/NUMA?

2. is my understanding above correct?

3. Is the current implementation of csched_load_balance() the result of evaluation even on Hyperthread/Multi-core/NUMA?

4. Do you have better idea for finding closest neighbor first?

5. Do you have an idea how to build the array, Neighbor[]?

 

Thank you.

-Caz Yokoyama, caz at caztech dot com. 503-804-1028(m).

 

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel