On Thu, 10 Nov 2011, Daniel Stenberg wrote:

> On Wed, 9 Nov 2011, Andrew Daviel wrote:
>
>> When transferring a large file over a high-latency (e.g. long physical
>> distance) high-bandwidth link, the transfer time is dominated by the
>> round-trip time for TCP handshakes.
>
> This is not true. The RTT is not really a factor for a single uni-direction
> stream in modern TCP since they introduced the ability to set large windows
> (>16 bit size), many years ago. If you believe otherwise, then please provide
> details.

I recall that the concept of "jumbo frames" was introduced, which I
presume is the same thing. I believe that that requires the large frames
to be supported end-to-end across all links, something that is not
guaranteed on the regular internet. Though the people I know that are
doing large long-haul transfers have a "light pipe" configured end-to-end
and can do that.

I haven't looked at the numbers, but I presume that as data size
and bandwidth increase the current large windows will become less
effective. I recall some kind of practical limit, but that may have been
for a current generation of backbone router.

> And for PUT it is not the same thing.
I haven't worked the numbers or tested that for real, only the principle.

> libcurl already supports HTTP/1.1 fully and allows applications to use
> libcurl to do many-connection transfers if they desire. Or are you talking
> about adding the ability to the curl tool?

I guess I was thinking about adding it to the curl tool. I'm just
floating balloons, really, so not too bothered if they get legitimately
shot down.

-- 
Andrew Daviel, TRIUMF, Canada
-------------------------------------------------------------------
List admin: http://cool.haxx.se/list/listinfo/curl-library
Etiquette:  http://curl.haxx.se/mail/etiquette.html