The conundrum facing many disaster tolerance and recovery designs is the tradeoff between loss of performance and the potential loss of data. On the one hand, it may not be desirable to slow application response time until it is assured that data will not be lost in the event of disaster. On the other hand, the prospect of data loss can be catastrophic for many companies and organizations. Unfortunately, there is not much of a middle ground in the design space and designers must choose one or the other.
The network-sync remote mirroring option potentially offers an improvement, providing performance of enterprise-level semi-synchronous remote mirroring solutions while increasing their data reliability guarantees. Like native semi-synchronous protocols, network-sync protocols simultaneously send each update to the remote mirror as the primary handles the update locally. Rather than waiting for an acknowledgment from the remote mirror, it delays only until it receives feedback from an underlying communication layer, acknowledging that data and repair packets have been placed on the external wide-area network. This minimizes the loss of data in the event of disaster. Applications requiring strong remote-sync guarantees can still wait for a remote acknowledgment, but for most purposes, network-sync represents an appealing new option. Our experiments show that SMFS, a remote mirroring solution that uses the network-sync option, exhibits performance that is independent of link-latency, in marked contrast to most existing technologies.