I got this from the Gilder thread . . .
I wanted to address (and perhaps even initiate a discussion) the technology of Mirror Image as seen by Neil Robertson, who is in charge of the adoption of emerging technolgies at Exodus. At last week's Exodus investors meeting, Neil discussed at length the inherent advantages of the MII topology - I will epitomize his remarks with a few of my own observations tossed in for fun.
MII is a content distribution platform on which EXDS has big plans for additional services. Right now the important part is completing the infrastructure (this is ahead of schedule by the way). Think of the different types of distributed architectures as a continuum - with one giant centralized server/cache farm containing all the internet content placed in the middle of the U.S. and the other extreme would be to place servers in every home and business throughout the globe. Each of these poles has benefits, what the giant cache lacks in speed of light improvement (for someone in germany) it gains in cache efficiency - this means that although it may take several 'hops' to get to the superserver, when your request arrives the information you're looking for will be there. The opposite is true for the cache on every table solution, your latency problem is solved, but your cache efficiency is abysmal. The only thing residing in your local cache is the information you just retrieved within the past week or so, and if the new information is not there, your speed of light latency benefits are usurped by the hops it will take to go out on the net to find it. So . . . .
When determining the right architecture for content distribution the variable of latency and cache efficiency are paramount. Neil also introduced the concept of management efficiency as an additional component to the equation. While these three variables are distinct, the end user experience is really defined by latency, or how fast the web page, or streaming video, etc... appears at my location. Perhaps this is the reason Akamai has focused on the highly distributed model, with over 4,000 servers sprinkled throughout the planet. What they lack in cache efficiency they try to make up with MIT-derived algorithms that incorporate intelligent routing to minimize hops if the content your looking for does not reside at your local server. A ha!!!
Cache efficiency is directly related to latency! What if someone said instead of placing all those thousands of servers everywhere, lets compromise and only build a network with 32 or so server/storage farms, but those nodes will be connected to all the major ISPs in the world. What we lose in a hop or 2 we gain in cache efficiency. Cache efficiency is a function of number of users and Mirror Image is 75%-85% efficient. Management efficieny is also a function of the number of servers, increase your nodes and managment efficiency decreases. Instead of a highly distributed network, Mirror Image found the sweet spot in a regionalized distribution model. Now I understand why Mary Collins asked Avi Friedman at the Telecosm why Akamai wouldn't be a customer of Mirror Image - Akamai needs the cache efficiency! (And incidentally the Content Bridge intiative is doing just that: Novell, Inktomi want to connect their cache appliances to the MII network - because it will improve the efficiency of BOTH solutions)
Other benefits of the MII CAP architecture include:
- decreased cache flush because of storage capacity - this improves performance and improves cache longevity (storewidth)
-no publishing process, slap a mi.com on the url and your content is on the system
-freshness - the more servers you have the more difficult it becomes to guarantee each server has updated content
-MII relationship with the ISP at their CAPs creates a postive feedback loop: as the instaspeed service improves efficiency the more content will be cached and the more efficient it becomes (again)
The MII-EXDS relationship is especially interesting because if you do happen to request information that is not at a CAP (a miss)- the EXDS network (and now GlobalCenters) is there with most of the top internet companies as customers - no need for MIT professors to figure out how to best route your request because its probably already on the network! Synergy!
And finally the regionalized topology allows you to roll out new services very easily with the amount of storage and processing power available at each CAP. A highly distributed architecure makes this difficult if you need to upgrade software or hardware at 4,000 sites to handle the new service.
Mirror Image is the only architecture to combine content distribution with caching while applying the concept of storewidth. Neil described the MII network as a platform on which the next phase of the internet will be built. I've decided its worth the risk to wait a few years (months?) to find out |
