On 2019-04-04 20:13, Venky Venkatesh wrote:
On a side note on the design, suppose you have 1 massive flow in terms of work per packet and is "long" lived. Then 1 core would be super busy while others would be idle. Then would it constantly consider migrating that flow since the busy-ness of the core would be above a threshold while others are idle? Again (after migrating) to figure out that the new core is super busy while others are idle and so on ...? A sort of thrashing effect.
When a port considers migration (to reduce its load), it consults the last 128 seen events dequeued on that port. If all those events originate from the same flow, no flow will be migrated.
