Video Upload Options

Virtual synchrony is an interprocess message passing (sometimes called ordered, reliable multicast) technology. Virtual synchrony systems allow programs running in a network to organize themselves into process groups, and to send messages to groups (as opposed to sending them to specific processes). Each message is delivered to all the group members, in the identical order, and this is true even when two messages are transmitted simultaneously by different senders. Application design and implementation is greatly simplified by this property: every group member sees the same events (group membership changes and incoming messages) in the same order. A virtually synchronous service is typically implemented using a style of programming called state machine replication, in which a service is first implemented using a single program that receives inputs from clients through some form of remote message passing infrastructure, then enters a new state and responds in a deterministic manner. The initial implementation is then transformed so that multiple instances of the program can be launched on different machines, using a virtually synchronous message passing system to replicate the incoming messages over the members. The replicas will see the same events in the same order, and are in the same states, hence they will make the same state transitions and remain in a consistent state. The replication of the service provides a form of fault-tolerance: if a replica fails (by crashing), the others remain and can continue to provide responses. Different members of the replica group can also be programmed to subdivide the workload, typically by using the group membership to determine their respective roles. This permits a group of N members to run as much as N times faster than a single member, or to handle N times as many requests, while continuing to offer fault-tolerance in the event of a crash. Virtual synchrony is distinguished from classical state machine replication because the model includes features whereby a programmer can request early (optimistic) delivery of messages, or relaxed forms of ordering. When used appropriately, these features can enable substantial speedup. However, the programmer needs to be sure that the relaxation of guarantees will not compromise correctness. For example, in a service that uses locking to protect concurrently updated data, the messaging system can be instructed to use an inexpensive form of message ordering, in which the messaging system respects the ordering in which individual senders send messages (FIFO guarantee) but does not attempt to impose an agreed order if messages are sent concurrently by different senders. Provided that the sender indeed held locks on the data, it can be shown that FIFO ordering suffices for correctness. The benefit is that FIFO ordering is much less costly to implement than total ordering for concurrent messages. To give another example, by delivering messages optimistically, virtual synchrony systems can outperform the Paxos that is normally required for implementation of state machine replication: Paxos normally requires a 2-phase protocol, whereas optimistic virtual synchrony protocols can deliver messages immediately upon their arrival. However, this could result in a violation of the safety property of the state machine replication model. To prevent such problems, the programmer who uses this feature is required to invoke a primitive called flush, which delays the caller until any optimistically delivered messages have reached all of the group members. Provided that the programmer understands this behavior and is careful to call flush before interacting with external clients or persistent storage, higher performance can be achieved without loss of safety. The flexibility associated with these limited forms of event reordering and optimistic early delivery permit virtual synchrony platforms to achieve extremely high data rates while still preserving very strong fault-tolerance and consistency guarantees.