A new technical report
For those of you who are interested in some of the design considerations of the multi-node data store in Darkstar, Tim Blackman and I have just published a Sun Labs technical report on the subject. These technical reports are the labs equivalent of the kinds of tech reports that you get from universities (or other industrial labs) and are meant to talk about the ideas behind the technology, rather than how to use the particular technology.
The Data Store is, as most Project Darkstar users know, one of the keys to multi-node. All data that is accessed by more than one task in a Darkstar based game needs to be a managed object, which means that it is going to be in the Data Store. This allows the Darkstar infrastructure to do contention management (which is necessary even on a single machine to make use of multiple cores), but it also allows the infrastructure to move tasks (or groups of tasks) from one machine to another without having to worry about not having access to the data, since the data store is shared by all nodes.
But to keep the latency at acceptably low levels, the data store has to avoid actually calling across the network of server machines whenever possible. This technical note shows how we (well, mostly, Tim) are trying to solve this problem. As I've said before, I think it is a pretty interesting approach, since it gives up one of the ACID (Atomic, Consistent, Isolated, Durable) properties usually associated with transactions to get the low latency.
Anyway, if you are interested, take a look. Lots of details are there.
