[ofa-general] Re: IPoIB path caching

[Jason Gunthorpe]

On Thu, Aug 02, 2007 at 08:24:10AM -0700, Sean Hefty wrote:

> >Does it mean that you will have to re-implement RMPP in a user space 
> >library or just the initiation of the query would be from user space?

I thought RMPP already was available to user space? In fact, I think
we have code someplace that uses it..

> RMPP would not need to be re-implemented, but I haven't come up with a 
> user space solution that I like either.  I'm not trying to design a full 
> blown distributed SA, nor do I want to try to standardize some new SA 
> protocol.  I only want to fix the SA PR scalability problems that we see 
> in practice.

Well, I still think the simplest thing is to make a new netlink
protocol to maintain a cache table in the kernel and then a simple
user space program that uses the user space RMPP interface (and trap
subscription..) to do GetTable queries and uses netlink to groom the
kernel cache. Basically exactly what you have now but seperated into
two halfs.

The main thing is to pick the kernel's table structure so that it can
be extended to support future things like multipath, QOS and routers..

I don't like the notion of communicating with the kernel using MADs,
it doesn't seem like it fits with the rest of the kernel networking
model. When/if a network component to SA distribution is built it
should be handled in user space and the kernel should be completely
unrelated..

Jason