Re: silly windows

> 
> It may be avoiding some of the REALLY silly windows like 100 byte 
> fragments of a 4K frame.  But at apparent_out_window/4, it can break a 4K 
> frame up into 4 framelets, and frequently breaks one up into 2 or 3.

So what is the right behavior?  That's the question, should we wait for
the window to open completely so we are sending the max we can in one 
chunk?  This works if the window is in fact larger than the frame,
but what if the fram is larger than the window, we have to fragment.

My off the cuff window/4 is to reduce "silliness" while still letting us
have a number of fragments in flight, think that it's advantageous to
keep the flow going.

I am almost tempted to do something like sending chunks of window size/4
if the frame size exceeds the window size, so that we can have frames
and seq's in flight and not be stuck in a cycle of filling the whole
window and having to wait for the seq in reqturn before sending more.

.
> 
> What would happen if I refused to accept anything less than 
> c->commit_frame->size?  Would that cause a stall of some sort?


I think so, certainly for c->commit_frame->size > c->apparent_out_window.

> 
> I note that c->apparent_out_window is only 4K, despite seemingly having
> set the window size to be 64K at both ends using bpc_set_channel_window()
> in pro_start_indication() and my _start()  function.  Am I setting the
> wrong thing, or in the wrong places?
> 

Possible that apparent_out_window is not getting maintained...  

> BTW, I only have one active channel right now, so various algorithms for 
> choosing the channel won't make much difference except to the extent they 
> disqualify our only channel from sending at the moment.
> 

True.

-bill


-------------------------------------------------------
This SF.Net email is sponsored by: IBM Linux Tutorials
Free Linux tutorial presented by Daniel Robbins, President and CEO of
GenToo technologies. Learn everything from fundamentals to system
administration.http://ads.osdn.com/?ad_id=1470&alloc_id=3638&op=click