These are actually quite interesting instance types. I can see the use for these.
How do Amazon handle the capacity planning internally? With a fixed resource allotment it seems pretty easy to figure out. Just divide the total capacity by the number of guests on an instance. With this setup presumably either, there are less guests on a host (to allow for more headroom), or if everyone bursts at the same time you don't actually get any additional capacity.
I suppose Amazon have a huge amount of data regarding how machines are actually used. They must have a fairly good idea of how much of the available power is used on average / other usage patterns that emerge. Would be very interesting data to look at.
It also begs another question. If you have seamless access to this additional capacity once you've earned the credits, could they introduce a system where you could have pay as you go access? I guess it's not that easy because at least the credit system puts a limit on how much each guest can burst.
Comments
These are actually quite interesting instance types. I can see the use for these.
How do Amazon handle the capacity planning internally? With a fixed resource allotment it seems pretty easy to figure out. Just divide the total capacity by the number of guests on an instance. With this setup presumably either, there are less guests on a host (to allow for more headroom), or if everyone bursts at the same time you don't actually get any additional capacity.
I suppose Amazon have a huge amount of data regarding how machines are actually used. They must have a fairly good idea of how much of the available power is used on average / other usage patterns that emerge. Would be very interesting data to look at.
It also begs another question. If you have seamless access to this additional capacity once you've earned the credits, could they introduce a system where you could have pay as you go access? I guess it's not that easy because at least the credit system puts a limit on how much each guest can burst.