AltraMayor
Where should we allocate mbuf memory pools?
Pull request #566 moved the allocation of transmitting and receiving queues of NICs to the NUMA node to which the underlying Ethernet device is connected instead of the NUMA node of the functional block responsible for the queue. There, it was a matter of requirements, but that requirement can be lifted by patching the bond driver. Is the new NUMA node choice faster?
Independent of the answer to the previous question, a possibly more important question is where should we allocate mbuf memory pools? Right now, Gatekeeper allocates these pools at the NUMA node where the responsible block runs. But couldn't having these pools at the same NUMA node of the NIC be a better option? This may avoid transmitting the whole packet between NUMA nodes by sending only the needed cache lines associated with the packets.
