Re: More on HTTP-NG timings

David Koblas (
Tue, 22 Nov 1994 09:50:58 -0800 (PST)

> HTTP 1.0 gave an average transfer time of 24.7 seconds. HTTP-NG 0.9 had an
> average time of around 5 seconds
> The same test, using a machine at Ventana Media (In NC, but routed via DC),
> gave an average time for HTTP-TOS of around 5.6 seconds, with HTTP-NG
> clocking around 1.5 seconds.

I have a few issues with HTTP-NG, on on a protocol level it didn't appear
as if there was a PROTOCOL-VERSION byte in the headers. Such things like
this are always important.

While it's great that you've created a session based control protocol,
so that you can take advantage of all of the years of development into
congestion control that TCP uses. There are a few small issues...

A) Over the Internet transfering 5 10MB files simultaniously is better
done with 5 .. 20 TCP connections. Becuase, if one TCP packet is lost
only 1 TCP connection drops into a timeout<->resend cycle.

B) If one were transmitting real multimedia object over the chanels,
having only one identified channel for data transmission make it harder
to do bandwidth reservation, since it is harder to determine if the
channel is a text file or a video stream.

I think the better thing to keep in mind with HTTP-NG is that is
a control channel that can transmit data (MGETs...) but has mechanisms
for allow easy creation of auxilariy data channels. Sounds simple,
but it should be supportable from a library level.

Also, how does your benchmarking of HTTP-NG compare with HTTP supporting
MGETs?  I would expect that they are within 10% of each other, since
most of the time saving is probably comming from the reduction in
TCP connection creation.