[Click] Userlevel Click and updated docs?

[Eddie Kohler]

Hi Armen,

On 05/24/2011 12:44 PM, Babikyan, Armen - 0663 - MITLL wrote:
> Hello,
>
> Is there an update to the ACM TOCS 2000 Click Modular Router document that
> includes Click developments accumulated since that document was published 11
> years ago?

Unfortunately no.  It's probably time to write and try to publish an actual 
paper with updates since then, but we haven't so far, preferring to work on 
features.  You can always look at the README and the NEWS file for high-level 
changes.

> For example, the wiki provides information about an experimental
> Multi-threaded Click, which contradicts some statements in the tocs00
> document.

Yeah, it's not really experimental any more; a number of people rely on it, 
both in the kernel and at user level.

> Also, the performance numbers cite a 700Mhz PIII, and hardware has
> certainly evolved since then.

It has indeed.  RouteBricks uses Click.

> The tocs00 document focuses on the in-kernel implementation, and suggests that
> use of the userlevel interface should be regulated to testing. I am curious to
> know whether the userlevel driver has improved significantly since that
> document was published, and whether "real" applications that use Click Modular
> Router do their business in kernel-space or in user-space.

It has improved.  However, it does not necessarily support the fastest 
transfer methods between kernels and user-level, or default to the fastest 
methods available.  On Linux, "CAPTURE PCAP" might currently be required for 
reasonable performance.  It would be wonderful to merge support for  PF_RING 
or another similar fast mechanism.  Patches welcome!
>
> This is significant for me because I am using Click as a harness for a Network
> Coding Engine (*) my colleagues and I have developed. Although our Network
> Coding Engine follows a very Click-like architecture, it works independently
> of Click and integrates with other environments like the OPNET Network
> Simulation environment. We haven't actively designed our software to run in
> kernel-space, and we'd like to know if it is inadvisable to generate
> performance numbers using Click's userlevel engine. If anyone has any
> docs/reports regarding experience/use/gotchas/etc of userlevel Click as a
> production platform, that would be great.

If you're not using floating point, your elemnets should likely work in either 
location.  Before generating userlevel Click numbers, DEFINITELY make sure 
that you get good baseline performance.  See above for advice.

>
> I'd be interested to hear your collective thoughts on the future directions
> for Click too.

Too general a question.  Click is still supported.  Click is still evolving.

Eddie


> Thanks!
>
> Armen
>
> (*) For more info on Network Coding, see
> http://en.wikipedia.org/wiki/Network_coding
>
>
>
> _______________________________________________
> click mailing list
> click at amsterdam.lcs.mit.edu
> https://amsterdam.lcs.mit.edu/mailman/listinfo/click