Compound Treatment of Chained Declustered Replicas Using a Parallel Btree for High Scalability and Availability

Abstract

Scalability and availability are key features of parallel database systems. To realize scalability, many dynamic load-balancing methods with data placement and parallel index structures on shared-nothing parallel infrastructure have been proposed. Data migration with rangepartitioned placement using a parallel Btree is one solution. The combination of range partitioning and chained declustered replicas provides high availability while preserving scalability. However, independent treatment of the primary and backup data in each node results in long failover times. We propose a novel method for compound treatment of chained declustered replicas using a parallel Btree, called the Fat-Btree. In the proposed method, the single Fat-Btree provides access paths to both primary and backup data in all processor elements, which greatly reduces failover time. Moreover, it enables dynamic load balancing without physical data migration, and improves memory space utilization for processing the index. Experiments using PostgreSQL on a 160-node PC cluster demonstrate the effect.

A fragmented data-declustering strategy for high skew tolerance and efficient failure recovery

“…Table 3 shows a balanced configuration after one node failure. Workload of the failed primary, C on N ode3, is shared by its replica fragments C1 and C2 on N ode (6,7) . 4 10 access to F and G on N ode (6,7) is redirected to N ode (9,10,11) .…”

“…Workload of the failed primary, C on N ode3, is shared by its replica fragments C1 and C2 on N ode (6,7) . 4 10 access to F and G on N ode (6,7) is redirected to N ode (9,10,11) . In turn, 1 10 access to I, K, and 3 10 access to J are resent to N ode (1,4,2) .…”

“…In our first experiments, we use the Compound Fat-Btree [7], which is a scalable distributed B-tree index for data location management in cluster systems. It indexes both primary and replica data within one index structure.…”

A Compound Parallel Btree for High Scalability and Availability on Chained Declustering Parallel Systems

“…The load can be balanced without data migration if the query frequency for data on PE 1 , PE 3 and PE 4 is α, while the query frequency for data on PE 2 is no more than 2α. We also assume that P i and B i (i ∈ [1,4]) in the table represent the same amounts of data. We assume that the query frequency is known merely for convenience, because many methods have been proposed for obtaining this information of any PE without any centered nodes, and [37] is one of them which is suitable for the Fat-Btree.…”

“…Here, scalability enables the database to grow while supporting an ever-increasing rate of throughput; availability enables the database to continue providing a full service without losing data when some parts of the system fail. Efficient accessing methods with skew-balancing ability and data replication with synchronization strategies are very important for achieving scalability and availability, respectively [1]. It has been proved that: consistency, availability and partition tolerance cannot be provided simultaneously in distributed systems, because providing guarantees for two properties will violate the guarantee for the third property [2].…”