Download Now Free registration required
An atomic multicast is a fundamental abstraction for building fault-tolerant systems. This paper assumes a system that is composed of data centers or groups that host many processes connected through high-end local links. It also suggests that a few groups exist that are interconnected with the use of high-latency communication links. According to another paper on atomic multicast protocols, no multicast protocol can deliver messages that have been addressed to multiple groups in one intergroup delay. This implies that in case a protocol is delivering a particular message, only the addressees of that message are involved in the protocol. This paper proposes use of a non-genuine multicast protocol so that it is able to deliver messages addressed to multiple groups in one intergroup delay. This paper also undertakes experimental comparisons against a latencyoptimal genuine protocol. These experiments show that the non-genuine protocol is able to provide enhanced performance in almost all the scenarios that are considered across the study. This paper also identifies a convoy effect in multicast algorithms because of which local messages are often likely to be delayed. That is, messages addressed to a single group by as much as the latency of global messages and messages addressed to multiple groups. This paper also proposes techniques that can be employed to reduce this effect. This paper also evaluates a latency-optimal protocol that is able to handle disasters.
- Format: PDF
- Size: 272.1 KB