[OpenAFS] 1.4.8, Rx Performance Improvements, and a Small Business Innovative Research grant

Ken Hornstein kenh@cmf.nrl.navy.mil
Fri, 03 Oct 2008 09:58:31 -0400

>With those observations... does rx-over-tcp look like a solution? On 
>the packet-transmission side probably, but the encapsulation very 
>likely still demands significant processing power. And running a 
>server with 10000 or 20000 TCP connections does not sound that obvious 

Here's what I can tell you, based on my experience:

- The encapsulation for rx-over-tcp wasn't, in my experience, a huge
  performance sink.  The trick was trying to create a protocol that
  didn't starve multiple writers who were trying to use the same
  connection.  I think I made some decent progress on this; I was able
  to convince myself that the problem wasn't insurmountable.

- Web servers seem to manage this sort of thing fine (admittedly, there
  are a lot of front-end devices they can use than OpenAFS cannot, but
  still, a lot of work has been done on handling tons of TCP connections).
  There are some obvious ways of dealing with this.  I think in practice
  it will be a non-issue.