Hello, developers. Attached initial benchmark of netchannel with memcpy() into kernelspace area, which could be mapped from userspace versus socket code using 1gbit link.
As you can see from the graph, netchannels outperforms sockets in speed,
but it's CPU usage is higher too.
It is possible, that it is the price, i.e. socket code would increase
it's CPU usage if it could increase it's processing speed.
Implementationis fairly ugly yet. It is required a some changes in
generic TCP state machine processing logic to clean things up,
so it would be performed not on top of sockets,
but using skbs from queues with appropriate parameters (timeout,
flags) provided either as new structure (so it would be embedded into
struct sock and netchannel) or as function parameters.
Netchannels currently use two queue dereferencings to work with socket's
queue processing:
- from netchannel's queue which is filled in interrupt
- from socket's queue which is filled in process context
which is a source of some speed problems too.
It still requires some thinking...
--
Evgeniy Polyakov
netchannel_speed.png
Description: PNG image
