> On 22 Jul 2019, at 20:11, Brian Dickson <brian.peter.dick...@gmail.com> wrote: > > Speaking mostly as a user, there is some really important information needed > for using VPP, which is not documented (at all, or at least anywhere that can > be found). > > I would like to make a request that someone take up this request and provide > documentation, and preferably add it (or a link to it) to the (overly sparse) > FAQ list. > > Here is what IMHO is necessary to document: > > RSS (and related techniques/technologies such as RFS, RPS, MSI-X, queue to > CPU affinity, etc). > > For example, if a NIC has been configured by way of ethtool before VPP takes > over the NIC, what happens to any RSS-type configurations, including "ntuple" > rules?
If you are talking about linux RSS config, it is gone. NIC is driven by user-mode driver which resets hardware on init. > If there is a plan for some of these things but no current code, this would > also be good to document. > > There are some configuration file elements (CPU and worker numbers/counts, > and dpdk device configuration for RX queues), but no indication how those > interact, and no apparent explicit control. Will be glad to explain if you provide specific questions. In general each NIC queue is handled by owning worker thread and you can see current assignment with "show interface rx-placement" command. If you turn on RSS, you will simply see more than one queue assigned (ideally) to different workers. "show runtime" will give you some idea how much traffic is handled by each worker. > > Ideally, it would be good to have the NIC itself handle the packet flow (to > specific queues) for ntuple rules, IF the NIC supports it, rather than having > to add explicit VPP Nodes for that. What my main concern is, is the ability > to dedicate CPU(s) for specific kinds of flow, especially if the flow > requires "punting" (or the vppsb/router or vppsb/netlink equivalents) for > host stack stuff, like handling BGP packets to a non-VPP BGP speaker such as > quagga or FRR or bird. You cannot rely on linux rss rules as they are reset. We have flow API which is able to stear specific traffic to specific worker thread but you will need to write code to do that. -- Damjan
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#13544): https://lists.fd.io/g/vpp-dev/message/13544 Mute This Topic: https://lists.fd.io/mt/32561940/21656 Group Owner: vpp-dev+ow...@lists.fd.io Unsubscribe: https://lists.fd.io/g/vpp-dev/unsub [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-
