On 10/25/22 03:02, Daisuke Matsuda (Fujitsu) wrote: > On Sat, Oct 22, 2022 5:01 AM Bob Pearson wrote: >> >> Make the internal interface to the task operations pluggable and >> add a new 'inline' type. > > I do not see why we need the new 'inline' type. It may be technically > possible to add it, but is there any situation where it is useful? > With the new mode, It will take longer for softirq (NET_RX_SOFTIRQ) > to complete its work, and that can cause a negative effect on the system > as a whole. > The one and only place where it makes sense in production is for the completer task for UD QPs. In the original rxe code it was called inline from the requester. Later that sort of got blurred and sometimes it was called inline and sometimes as a tasklet. For benchmarking and testing it is useful to have other options to check. In my world, HPC, storage use cases are typically all consuming when active and we will try anything to see what gives the best performance and not worry too much about the effect on other jobs. I am also interested in trying threaded interrupts which are described as another alternative to tasklets. Bob
