Hi, This concerns eventdev being used in a DPDK multi-process mode wherein the PRIMARY process sets up the device, ports, queues and linkages and the SECONDARY processes are the real workers to which the events are load balanced to via the queues.
My question (for both the sw evdev PMD and the DSW evdev PMD) is what is the recommended handling when one of the SECONDARY processes dies? In answering this you can assume that the dead process will be restarted in a few seconds (within 10 seconds): 1. Is it worthwhile unlinking that process from the queues it is linked to? 2. If so, do these PMDs support such capabilities? Additionally, what is to be done with (i.e. with respect to telling the eventdev) the events queued to the concerned core but not dequeued AND the burst dequeued held by the process at the time of death? 3. If not, then if there is continuous traffic bound to that process (for reasons e.g. scheduling algorithm of the PMD, flowid state while dying etc.), will the device eventually get backed up due to max-inflight? If so, what is the recommended remedy? 4. For DSW there is the additional aspect of ongoing/future migrations -- what is the design recommendation to compensate for that during a process crash If your answer is sensitive to the restart delay, pls explain for the different delay ranges. Thanks -Venky
