Here's a great paper from Bolosky and others at Microsoft that demonstrates that Paxos can indeed be used as a high-volume replication protocol. They argue that a Paxos-based storage service can provide performance close to what the underlying hardware can support. You don't have to resort to simple primary-backup schemes which make it difficult to deal with arbitrary machine restarts. Also, you don't have to give up sequential consistency for performance and deal with the complications of eventual consistency. The crux of their argument is: for a system that is in a single datacenter, and needs to use commodity networking and disks, the Paxos implementation will certainly not be the bottleneck.
They implemented this in the context of a storage system called Gaios. The paper has plenty of implementation details and performance results. They even ran an OLTP benchmark on SQL Server configured to use Gaios storage. Neat stuff!
Spinnaker exploits the same ideas as Gaios, but the exposes a user-programmable key-value store API instead of building scale-out storage. The results from Gaios independently verify the arguments we tried to make in the Spinnaker paper -- you can use a consensus algorithm for data replication in a scale-out system without sacrificing performance.
They implemented this in the context of a storage system called Gaios. The paper has plenty of implementation details and performance results. They even ran an OLTP benchmark on SQL Server configured to use Gaios storage. Neat stuff!
Spinnaker exploits the same ideas as Gaios, but the exposes a user-programmable key-value store API instead of building scale-out storage. The results from Gaios independently verify the arguments we tried to make in the Spinnaker paper -- you can use a consensus algorithm for data replication in a scale-out system without sacrificing performance.
No comments:
Post a Comment