February 10, 2007


A few years ago an old acquaintance called me up and said that he had gone to work at Netli.  I went to visit and found the people at Netli to be imaginative and their ideas interesting.

I recently heard about the acquisition of Netli by Akamai - I hope my acquaintance will be making a goodly amount of money.

Part of what Netli does is to replace TCP with a proprietary protocol between what are, effectively, HTTP proxies.  This kind of innovative deployment of ideas is why the end-to-end principle and The First Law of the Internet are so important.

However, there are concerns.  Over the years the TCP protocol has been refined through the addition of congestion detection and congestion avoidance mechanisms.  A properly implemented TCP engine will degrade the performance of its individual connections in order to protect the internet as a whole.

TCP implementations that don't follow these congestion rules can be considered as taking unfair advantage of the good graces of those TCP implementations that do.

I do not know any details of Netli's protocol so the concern that I express below may or may not have grounding in fact.  But even if Netli has done it right, others may not.

The internet works because it is a vast statistical multiplexor.  As such it is sensitive to saturation by streams that do not notice congestion, much less back off when congestion occurs.  This is not a conjectural concern; when I was inventing and implementing IP/TV a decade ago we saw the negative effects of heavy streams of video RTP/RTCP (UDP) packets that kept coming at full rate even when a router or link had become saturated.

A similar situation existed long ago on coax-cable based Ethernets - some Ethernet controller chips (such as some from Sun Microsystems) seemed to have more aggressive timers than those used in PC's.  The effect was that an ethernet that worked fine among PC's suddenly became unusable for those PC's, but not for the Suns, when one or more Sun machines was attached

My concern is that those, like Netli, who design TCP alternatives ensure that their protocols play fair and don't try to take advantage of protocols that do back off in the face of congestion.

Posted by karl at February 10, 2007 2:05 AM