Perf: Profile-based buffer options generation should be based on average message size estimation #132
Description
Context
In #131, we introduced Profile, which is an enum representing different performance profiles ((Throughput, Latency, Balanced)). These performance profiles are now just used to set buffer size and buffer linger options to some reasonable values.
To make them maximally useful though, they should take into account an optional estimation of the message side that the caller can provide, primarily to tune the buffer size. If your average message size is many multiples greater than the Throughput profile buffer size, the socket driver will still have to make a lot of syscalls just to write your message to the underlying socket. We should try to minimize these syscalls because they're expensive.
So, a buffer size tuned to message size could be the highest multiple of 2 that is still smaller than the estimated message size.
More context:
#131 (comment)