On Wed, 2014-08-27 at 15:17 -0400, Ted Ross wrote:
> 
> On 08/27/2014 03:05 PM, Andrew Stitcher wrote:
> > On Wed, 2014-08-27 at 14:55 -0400, Andrew Stitcher wrote:
> >> ...
> >> Of this only applies under the exact circumstances of these tests, until
> >> we have more data points. I'd suspect it must also depend on the message
> >> size too.
> >>
> > 
> > Just to reply to myself...
> > 
> > Keeping the credit window at around the MTU size is going to give you
> > the most efficient networking transfers.
> 
> I'm not sure I'm following you here.  Credits are in units of messages
> whereas MTU is in octets.
> 
> To avoid bad interactions with the Nagle algorithm in TCP, the credit
> window multiplied by message size should exceed MTU (Maybe that was the
> point you were originally making).

Not clear enough - (fuzzy thinking today) - what I'm really trying to
say is that for efficient transfers you want the packets to be full, so
the credit should be high enough so that credit * msg size > MTU. But,
as you say, the latency is just as important too. s\So actually I should
have really been pointing out that the aim is to fill the
bandwidth-delay product, so that you send new messages before the
messages in the pipeline run out.

Andrew


Reply via email to