Hi All, As per my current understanding of the code for ARP learning:
1. ARP is learned on a VPP worker 2. Send to main thread via rpc, by adding to the rcp queue 3. main thread picks up this from rpc queue 4. main thread takes thread barrier lock and updates the ARP table In the step4, we are taking a thread barrier lock for ARP learing ( might be due to the ARP module being non-thread-safe). We are observing that this is resulting into worker threads being locked up for few milli/microseconds, causing some tail drops. Since, ARP learning is quite a common activity in the network, isnt this causing issue for VPP in general. We are seeing tail drops in VPP workers during ARP spikes. Any ideas on how to handle this (or) making ARP learning as lock-less, so that it can be done from any worker ? -- Thanks & Regards, Murthy
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#20017): https://lists.fd.io/g/vpp-dev/message/20017 Mute This Topic: https://lists.fd.io/mt/85132512/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-