On Fri, Aug 30, 2013 at 06:00:50PM +0000, Terry Gilsenan wrote:
> -----Original Message-----
> From: owner-postfix-us...@postfix.org 
> [mailto:owner-postfix-us...@postfix.org] On Behalf Of Robert Sander
> Sent: Saturday, 31 August 2013 3:47 AM
> To: postfix-users@postfix.org
> Subject: Re: Re-inventing TCP (was: newbie check..)
> 
> >Am 30.08.2013 19:36, schrieb Terry Gilsenan:
> 
> >> The killer on high latency links is the tcp-window and the continual 
> >> wait for ack. With links above 1000ms this compounded delay reduces 
> >> the available bandwidth to a very small percentage of the interface 
> >> speed (eg:256kbps on a 2mbps link). Without this I have seen UDP data 
> >> streams that will approach the actual interface speed. (eg: 1.8mbps on 
> >> a 2mbps link)
> >
> >That may be true. But SMTP is an non-interactive, non-realtime application 
> >that does not need high bandwidth. Keep the servers talking to each other 
> >until the >email is reliably transferred over TCP.
> 
> Yeah, you know that, and I know that, but users that are connected with a 
> 1mbps vsat link (with varying latency between 900ms and 1300ms) who then 
> attach a 20mb document to an email and click send, cannot understand why it 
> takes 45 minutes to deliver.
> 
> That same 20mb over UDP will only take a couple of minutes, and in reality 
> does not use any more bandwidth. 
> 
> When you extrapolate that by say 50 users sharing that 1mbps vsat link, and 
> all of them using SMTP to send attachments of various sizes, saying be 
> patient simply doesn't work because the tcp-window makes the link very 
> inefficient, and a given the cost of vsat bandwidth (in the region of 
> US$5k/mbps/month), just buying more is not a sensible option either. (not to 
> mention the infrastructure cost escalation with higher upload speeds)
> 
> I think the implementation of an additional feature that will open a UDP 
> stream for bulk data in parallel to the TCP connection through which the SMTP 
> transaction is happening is workable.
> 
> Other protocols do a similar thing, so it is more along the path of embrace 
> and modify than to re-invent the wheel.
> 
> T

Hi Terry,

Have you tried using a tcp-over-udp tunnel transport?

Regards,
Ken

Reply via email to