Application Architectures - The Practice of Cloud System Administration - page 83

machines that cache large amounts of content. For example, they may store the 1000 most
popular videos being viewed at the time ( Figure 4.5 ) .
Figure 4.5: A private network backbone connecting many datacenters (DCs) and points of
presence (POPs) on the Internet
Thus an internal network connects datacenters, POPs, and satellites.
Global Load Balancing and POPs
Thisbringsupaninteresting question: shouldtheGLBdirect traffictothenearest frontend
or to the nearest datacenter that has application servers for that particular service? For ex-
ample, imagine (and these numbers are completely fictitious) Google had 20 POPs, satel-
lites, and datacenters but Google Maps was served out of only 5 of the datacenters. There
are two paths to get to a server that provides the Google Maps service.
Thefirstpathistogotothenearest frontend.Whether thatfrontendisinadatacenter or
satellitedoesn'tmatter.There'sagoodchanceitwon'tbeinoneofthefivedatacentersthat
host Google Maps, so the frontend will then talk across the private backbone to the nearest
datacenter that does host Maps. This solution will be very fast because Google has total
controloveritsprivatebackbone,whichcanbeengineeredtoprovidetheexactlatencyand
bandwidth required. There are no other customers on the backbone that could hog band-