Re: [PATCH] new UDPCP Communication Protocol

[Stefani Seibold <stefani@seibold.net>]

Am Montag, den 03.01.2011, 11:39 +0100 schrieb Eric Dumazet:
> Le lundi 03 janvier 2011 à 10:54 +0100, Stefani Seibold a écrit :
> 
> > How can you do a routing, how can you determinate the MTU of the route.
> > This are basics. Look into other code how this things will be handled is
> > in my opinion the right way, since there a no function provide to do
> > this.
> > 
> 
> Hmm, how user land can perform this task then ?
> 

Userspace is much more complicate and more overhead than kernel space.
The UDPCP implementation in userspace is about the factor 10 slower.

> Is there an open source implementation of UDPCP ?
> 

I don't know any. These is the first one.

> What are its problems ? You say its dog slow, I really wonder why.
> UDP stack is pretty scalable these days, yet some improvements are
> possible.
> 

UDP is fast... but UDPCP depends extremely on latency due the missing of
sliding windows. 

> Why not adding generic helpers if you believe you miss some
> infrastructure ? This could benefit to other 'stacks' as well.
> 

Maybe i don't have the knowledge, maybe i don't have the time. Getting
in new API functions into LINUX is much more complicate than getting new
driver into LINUX. I know what i am talk, it takes me one year to the
new kfifo API (kfifo.c, kfifo.h) into the kernel.

> > Otherwise you can say the same about all the filesystem or PCI
> > drvivers , which do also a lot in the same way. But since this is the
> > way to do it, it is the right way.
> > 
> 
> These drivers are here because of high performance on top of high
> performance specs.
> 
> While UDPCP is only a layer above UDP. If the problem comes from UDP
> being too slow, it'll be slow too.
> 

Because of latency. Handling the UDPCP into the data_read() bh function
is much faster:
- No context switch
- Assembly Multi-Fragment Message is very efficient using skb buffer
chaining.
- Immediately handling an ack or data message save a lot of latency

Implementing it in User Space is to slow, due the context switches. Also
the sunrpc approach is not faster due the using of kernel threads which
are not better than user space (okay, a little bit because not switching
the MMU).

The implementation is clean. I did fix all issues what i was asked for.
The protocol has now absolut no side effects. So i ask again for merge
into linux-next.

- Stefani