Re-inventing TCP (was: newbie check..)
- Terry Gilsenan:
> I have done testing with file transfers using both TCP and UDP,Would not the same result be achieved by disabling TCP congestion
control? I am not implying that doing so is a good idea.
- On 8/30/2013 10:12 AM, Terry Gilsenan wrote:
> I am not talking about implementing SMTP on UDP, I am taking about the possibility of adding a side-channel for bulk data that would use UDP.I'm really surprised nobody has mentioned this yet. It seems there's a
far simpler solution to the described high latency low bandwidth
attachment transfer problem: TCP send/receive buffer size.
If all receiving systems advertised a sufficiently large receive buffer
and clients could match it, then this latency/bandwidth problem pretty
much evaporates. A 1MB buffer allows a 50MB attachment to be sent
requiring only 50 ACKs. Assuming the path is fairly reliable, which has
been your assumption in your UDP argument, then this is a far less
invasive solution with similar performance.
IIRC, not may years ago, a team comprised of personnel at Argonne and
CERN achieved sustained saturation of a 1 Gbps link between Chicago and
Geneva, using parallel FTP and an insanely large TCP buffer, like in the
multiple GB range. I'll have to dig it up.
This demonstrates that high bandwidth TCP over high latency links is