0807

[HigashikataZhangsuke]

1.Fix IsoInvoker -> 10+h. See 0806

What kind of resource usage do you want?

[HigashikataZhangsuke]

How do we modify multiple functions concurrently? Or just use serial. Check Overhead. If not too much, serial is ok, simple

[HigashikataZhangsuke]

How to use MBA here:

1. Who will need MBA: same M type function and Share part -> you only need 5+1 = 6 COS <8
2. The general idea: limit the one has biggest BW requirement. Previous exp get the result that if we apply MBA and CAT under not overloaded conditioin is working. For overloaded, need to (1)use exp to show this works (2)overloaded is not the common case -> You assgined to much BW for the function you do not mean to.

The Memory Bandwidth Allocation (MBA) feature provides indirect and approximate control over memory bandwidth available per-core. Therefore, I think the amount you allocated has problem. See https://github.com/intel/intel-cmt-cat/issues/277. Like if I want to apply 4000 for all, then, it means percore 4000, but not overall 4000. So that means, you MBA allocation's value should be aligned with your profiling data, but you need to change the CPU assigned to this COS.

Another thing need to findout is: what if you applied more than the overall BW system could provide.

Maybe Applying CAT is not good. only 11ways, and does not affect much.

According to the test results, MBA is overall but not per core

Over provide BW for functions have nothing effect on throughput.

[HigashikataZhangsuke]

Invoker still has parts did not finished

[HigashikataZhangsuke]

No You are correct, it based on per-core manner. You need to figure out the behind reason for your test.