Greetings
Our app has several multicast udp/ip listen and several tcp/ip connections simultaneosly opened. (14 or so UDP and 6 or so TCP sockets)
It uses non-blocking select/recv polls in a single thread or several threads (1 thread per cpu). Currently we use kernel stack with interrupt moderation turned off and 100% cpu utilization for polling.
We look forward to gain low-latency benefits from using vma.
We can rewrite our app to blocking sockets, if it gives any benefits
We don't understand the vma architecture to properly configure offload for lowest latency possible.
How to poll offloaded sockets, i.e blocking or nonblocking?
From the vma manual we figured out that it has one or more inner threads. Do they still perform when we poll offloaded sockets 100% cpu time?
Are there interrupt affinity considerations for a single processor system?