Re: [libdc] Reducing latency with a packet ring buffer?

[Damien D. <dd...@do...>]

On Sat, 2008-12-06 at 17:15 -0800, David Moore wrote:
> Yes, this would certainly be possible, at least on Juju and Mac OS.
> 
> In fact, in some sense we are already using a packet ring buffer on 
> those platforms.  The data is put into the framebuffer by the DMA 
> program and will be visible prior to dequeue_frame() being called.

So the only thing we need is to know exactly how many valid bytes we
have?

> We could add some new APIs that exist in addition to the existing APIs 
> that would be for "advanced users only" and would let you "peek" at the 
> contents of the buffer prior to it being filled completely.
>
> The main tradeoff is that additional interrupts would be generated. 
> Right now, there is only one interrupt per frame, which is very 
> efficient from the kernel's point of view.  If we added additional 
> interrupts, that would consume additional CPU.  Of course, that might be 
> enabled at run-time only if a user needs it.

Do we necessarily get extra interrupts if we just have to update a
counter?  If the interrupts are a problem we could maybe finetune their
frequency (e.g. update the valid byte counter every X packets) 

Yes, I don't know much about the kernel ;)

-- 
Damien 高原 Douxchamps    http://damien.douxchamps.net/