Economical Fault-Tolerant Networks

If unchecked, any alien machine can generate an election call and the servers will move into an unwanted election state. For this reason, security and integrity of the signals exchanged between servers is highly stressed. We have implemented three levels of security for safe operation.

For our implementation, we empirically found that weekly synchronization of large data suffices, while the password databases are replicated on every election. Other scenarios may require a more frequent synchronization.

The level of reliability and fault tolerance of the cluster increases in proportion to the size of the cluster. Increasing the number of servers increases the maintenance time for synchronization and unduly lengthens the election process. In our situation, we determined from experimenting that three to four servers are enough to guarantee a practical working solution.

On average, the slave servers are negligibly loaded, while the master servers are not loaded for more than 0.1% of the CPU usage. Network load is also very low, except during heavy synchronization, which is therefore run as a scheduled process.

Our test bed for this implementation is the Digital Computer Laboratory, UET, Lahore, Pakistan. The lab consists of 10 Pentium-based servers and 60 diskless workstations connected by 10MBps Ethernet.

There is a need for development and implementation of techniques that provide for an immediate synchronization. One method could be that whenever a file is updated, all the servers update their versions of the file. In this way, all data on all servers at all times is perfectly synchronized, thus eliminating heavy network/server loads during scheduled synchronization.

This technique is a practical and feasible implementation of fault tolerance for low-budget LANs running open-source operating systems, such as in developing countries and resource-scarce academic institutions where expensive commercial solutions are just that—expensive.

Acknowledgements

Resources