Re: Improve performance between Network driver domain to guest OS

[Wei Liu <wei.liu2@citrix.com>]

Please don't drop xen-devel.

On Sun, Oct 13, 2013 at 12:12:26PM +0530, shikha thakur wrote:
> Hello,
> I apologized for the late reply.
> 
> >>What's your goal? Higher throughput? Lower latency? Do you have any
> merit in mind?
> 
> *Me: Basically higher throughput is my Aim.
>        **To increase the throughput, the paper i am referencing  analyzed
> the Xen hypervisor and got the memory latency problem.It has been overcome
> by packet aggregation.Thus the higher throughput is achieved.With the
> aggregation mechanism ,the throughput almost increases linearly  with the
> number of VM.(As Shown in Figure below).However, It introduced delay &
> jitter , hence i am trying to either eliminate delay & jitter or find new
> way of increasing throughput.
> 
> *
> 

I skimmed the paper. "Memory latency" seems to refer to the latency
introduced by memory copy. And the solution in that paper might not work
quite well in the scenario when a VM is expecting low latency with small
packets.

> 
> >>It depends. AIUI packet aggregation can introduce latency too. The only
> way to prove whether it works is to implement it.
> 
>  *Me:: It has been experimented and it is introducing delay and jitter.
> *
> *First thing i wanna know, Whether the  maximize delay and jitter is really
> a problem for higher throughput?

Sorry I don't know. you'd better think of the system as a whole.
Eliminating one single bottleneck might actually decrease performance.

> *
> *
> If yes, there are two method :-
> 1) Packet dimensioning tool which dynamically tune the packet
> aggregation.(Refer attached pdf)**
> 2) What i am planning to do is to imply adaptive jitter buffering algorithm
> to dynamically tune the packet aggregation*.
> 
> *If not of much trouble,Could anyone guide me which one of the above 2
> method is better for minimizing delay and jitter.*
> *
> 

Sorry I cannot give you a definitive answer, that's beyond my expertise.
I think the fun of doing research lies in the uncertainty and the thrill
of solving problems. If there's definitive answer to this question,
somebody would have done it already, right?

The only viable way is to hack a prototype IMHO.

Wei.