George Gilder is a widely published columnist and pundit. His thoughts about man’s technological future are widely respected. Lately, he has been evangelizing a vision of an all-optical Internet, plentiful bandwidth and the demise of many traditional Internet technologies like Cisco routers. Some of this vision is underpinned with assumptions that are not applicable and as clarity in assumptions is critical for any future vision, I’d like to use this forum to work through these.
First of all let me quote Gilder directly so that my jumble of opinions do not dilute his message:
“… the Internet is a computer on a planet. Like a computer on a motherboard, it faces sever problems of memory access. The Internet communications depend on ingenious hierarchical memory management, analogous to a computer’s registers, buffers and latches, its three tiers of speculative caches, its bulk troves of archives, its garbage management systems to filter and weed out redundant or dated data, and it direct-memory access controllers to bypass congested nodes.
In a world of bandwidth abundance, an ever-increasing share of roundtrip delay for a message is attributable to speed of light latency. No matter how capacious the transmission pipes, how large in numbers of bits per second the data stream, the first bit in the message cannot move from source to terminal any faster than light speed allows, plus the time waiting in queues and buffers at all the switches or other nodes along the way.”
He then makes the case for strong caching and dispersed assets … here are some of the supporting assertions:
“Even with no hops or other delays, the light-speed limit alone means that Internet users outside the North American continent are at lest 200 milliseconds away from the vast majority of websites.” (80% of web hosts are in continental USA)
“To fetch a web object using the Internet protocol – whether a frame, image, logo, or banner – takes two to seven round trips between the end user and the Web server. With each page comprising as many as twenty-five objects, those round-trip-speed-of-light milliseconds keep adding up even for entirely static material.”
Perhaps a little table will help us keep track of what physics foundation Gilder is operating from.
Description value units
speed of light 300000 km/sec
diameter of earth 12500 kilometers
circumference of earth 40000 kilometers
max intercontinent hop 30000 kilometers
round trip time 200 milliseconds
typ intercontinent hop 12000 kilometers
round trip time 80 milliseconds
typical US only hop 2500 kilometers
round trip time 17 milliseconds
So, his assertion of 200 ms for the worst case hop just based on speed of light arguments is well justified. The figure hinges on the following underlying assumptions worst-case distance is:
- 50% of earth’s circumference
- 50% longer than even this due to routing inefficiencies
- really 2x this because a web page fetch involves a request and an answer so the distance is traversed twice.
This worst case figure is really exaggerated for the point of his article … a more typical intercontinental distance is likely to be 12000 km and then the time drops to 80 ms. And for normal intra-continental distances of 3000 km this is then 17 ms. Actually most traffic will be of this latter category since each citizen/user will most of the time be fetching information localized for their country.
But then his assertion that this is the root of all our problems does not wash. He assumes that web pages will always be transmitted as individually requested objects on a page. This assumption is weak … a much simpler solution is for HTTP and its cousins to migrate towards a single transmission request for all information and for the sender to gather together the information and send it just one time as a single group or stream. No doubt this is better for the server and the Internet infrastructure to see a larger single flow instead of so many little (25) pieces. Again, it is the multiplication of so many separate requests to assemble a single web page that causes the problem.
In a sense, such progress assumes that Internet protocols will continue to evolve … sounds reasonable. Moreover, once the streaming has started the end user will not notice the effects of distance provided there is sufficient bandwidth between the sites and buffer at the receive end to smooth out any potholes in the transmission
Which is the better approach?
Reasonably local/short distance widely dispersed caching or more aggregated page serving?
Let’s use Gilder’s assumption of 25 objects per web page for our starting place. We shall assume the cache resides at an effective distance of 250 km for a round trip time of 1.7 ms and therefore 25 objects will take 43 ms. This is substantially slower than the direct sending of the page over typical intra-continental distances of 2500 km but sent as a single object (17 ms).
The point of this is that direct bundled sending is substantially better than local caching of pages like Akamai or Inktomi can provide.
As our agents and web servers get smarter, we will find that their ability to coordinate the information flow to us so as to minimize our waits and keep their responses more in line with our normal cadence of work will remain acceptable.
So where is the problem of Internet scaling? It is primarily in the core of the network where the global computer must continue to scale to packet per second forwarding levels measured in terms of 1012 and then 1015 packet per second over the next few years. Today the biggest, baddest layer 3 switches or routers have ability to do about 108 packets per second. Collections of these cannot do any better than this because they are ultimately placed into chains and the chain is only as good as its weakest link.
The problem with Internet scaling is not with typical web pages or the migration of television or even telephones to the web; it is with distributed computing. As we seek to create larger clusters of computing, especially as clusters are formed and broken down dynamically as their user bases require this distance and light travel time strongly affect the ability to do cluster computing.
Interestingly enough the problem in trying to create a world wide computer system (like that envisioned in the movie, The Matrix) is that the messages will only propagate at the speed of light and the network in between cannot beat this limitation. Responses that require inputs or outputs from disparate points in the system will not respond in computer time but in earth time … milliseconds not nanoseconds.
(this essay was originally written in March, 2000. fortunately the physics has not changed!)
Posted in Optical Technology