From Mike Axelrod - Google
Google Global Cache (GGC) allows a large portion of requests to be served from a small node located inside the service provider’s network serving regional users
Significant performance improvements for proxied connections over high latency links. (20%)
Based on user location and the type of content requested Google DNS will direct user to the best Google node
If given node already has the requested content in its local cache, it will serve the content directly to the end user.
When content is served from the node end user experiences improved performance and the service provider does not incur the expense of carrying the request and service traffic across their peering and/or transit links.
If the content is not already stored in the node, it will retrieve it from Google and store it for future requests.
How it works
User resolves content_host.google.com. (example only)
If the ISP DNS doesn’t already know the IP address, it queries Google DNS for the IP address of content_host.google.com.
Google DNS server knows that this ISP has a Google Global Cache node that should be able to service the request, so it replies with that IP address.
IP address is returned to the user.
The user requests the desired content from the IP address, which happens to be a GGC node on their ISP’s network instead of a server on the Google network.
If the content is cacheable and not already on the GGC node, it requests the content from Google on behalf of the user and caches it for future requests.
Once the GGC node has the content, it can serve it to the requesting user as well as subsequent users looking for the same content.
Key is to select Google server in close proximity to the requesting client
Most CDN systems today use the DNS to make such server selection decisions. However, DNS provides only the IP address of the client's local DNS server to the CDN, and not the user's IP address.
Implicit assumption that clients are close to their local DNS servers could lead to suboptimal node selections
The only way to improve the situation is to ensure that users point to right resolvers
IP-2-GEO techniques used to make sure that DNS IP and user IP map to proper geography where node is located
GGC gives control to ISPs by collecting IP prefixes using BGP from ISPs hosting GGC - to determine which users should be served by a given node.
Proxy mode: GGC pre-establishes connections to Google back ends, terminates user TCP session locally and is capable of caching static content.
Significant speed improvements over the high latency links due to the elimination of at least 3 rrt (syn/syn-ack/ack) and reducing server time-outs.
GGC runs on rack mountable servers 4-6 in each cluster
Each server provides back-up to others in a cluster
Deployment in ISP networks that have substantial traffic to Google and are interested in saving BW
o rack space 8-12 RU
o power ~ at 1800-2700W
o IPs and Ethernet ports