The paper describes a technique to correct crash faults in large data structures hosted on distributed servers, based on the concept of fused backups. The prevalent solution to this problem is replication. To correct f crash faults among n distinct data structures, replication requires nf additional replicas. If each of the primaries contains O(m) nodes of O(s) size each, this translates to O(nmsf) total backup space. Our technique uses a combination of erasure correcting codes and selective replication to correct f crash faults using just f additional backups consuming O(msf) total backup space, while incurring minimal overhead during normal operation. Since the data is maintained in the coded form, recovery is costly as compared to replication. However, in a system with infrequent faults, the savings in space outweighs the cost of recovery. We explore the theory and algorithms for these fused backups and provide a library of such backups for all the data structures in the Java 6 Collection framework. Our experimental evaluation confirms that fused backups are space-efficient as compared to replication (almost n times), while they cause very little overhead for updates. Many real world distributed systems such as Amazon's Dynamo data store use replication to achieve reliability. An alternate, fusion-based design can result in significant savings in space as well as other resources such as power.
展开▼
