[Ns-developers] GSOC & Optical Burst Switching Network patch port.

[Tom Henderson]

rAjEs wrote:
> Hi 
> 
> Some confusions about Linux kernel network stack integration project 
> idea. I have copy the project idea in one paragraph and my thoughts in 
> the other paragraph with dash lines. I need to rectify my understanding 
> so that i can modify my applications.
> 
> The goal here is to be able to run a minimally modified (ideally, 
> unmodified) linux kernel network stack (initially, ipv4/tcp) within our 
> network simulator. There are two distinct components to this project:
>  
> ---It means that are we modifying the code of kernel (like 
> para-virtualization concept). Since the NetDevice Class is the class 
> which defines the IP and ARP layer and it also manages  the instances of 
> a network driver(Here network driver means the user-space network driver!!).

rajes,
the NetDevice class does not define the IP and ARP layers; they sit 
above this class.

> 
> Implement a set of glue layers to provide a 'kernel' environment for 
> kernel code. This will require: a re-implementation of the basic kernel 
> library functions (memory allocation, timers, etc.), the implementation 
> of a fake linux network driver to act as a bridge between the ns-3 
> NetDevice class and the linux struct net_device, and, the implementation 
> of an ns-3 fake socket to act as a bridge between ns-3 applications and 
> linux kernel sockets.
> 
> --- what do you mean "kernel environment for kernel code". 

This means that we want to provide an environment that matches the 
kernel environment as much as possible, to avoid making unnecessary 
changes to the kernel code to get it to run within ns-3.

As far as i
> understand we are modifying the kernel library. So then our kernel will 
> be custom kernel(confused). This fake driver you mean reside on user 
> space or kernel space . Since as i know Class NetDevice is analogous to 
> the interface above the dev_queue_xmit() of linux kernel network code. 
> This function uses the sk_buff (Data movement for sockets takes place 
> using a core structure called the socket buffer). And can you elaborate 
> about the ns-3 fake sockets.
> 
> ---- Till know i am understanding that this linux kernel network 
> integration can be implemented into linux kernel or user-space driver. 
> And where are we planning to implement ?. If we are implementing the 
> fake driver in kernel can we use NAPI apis for the network.

The modified kernel code would run in the user-space ns-3 process.

> 
> Re-use the network namespace code to allow instanciating multiple 
> network stacks within a single address space. This area of the kernel 
> code is still undergoing heavy development so, we expect that a lot of 
> debugging will have to happen, as well, as, potentially, cleaning up 
> some rought edges of the kernel implementation.
> 
> ----- As stated the kernel code is undergoing heavy development in this 
> area. What are we planning to do here ?? Just the debugging or cleaning 
> the code or stick with the one linux kernel version and clean it .

IMO, it would be useful to have a technique that was able to track the 
Linux kernel tree rather than freeze at a particular version.  But that 
would be much more challenging since the kernel APIs are not stable.

- Tom