The dissertation focuses on fast committing protocol and higly efficient recovery schems for parallel main memory database on clusters, including: fast transaction committing protocol and logging, checkpointing, and recovery of the parallel main memory database.The dissertation has enhanced traditional one phase committing protocol, and propose using parallel log writing to fully utilize the IO bandwidth to accelerate transaction committing, thus avoided log accumulation in update intensive applications. Traditional two phase locking leads to lock conficts between read only transations and update transactions, which lowers down system throughput. A novel transaction schedule protocol is proposed in this dissertation, transient versioning is combined with locking to support non blocking reading, and avoid conflicts between reader and writers.The consitent checkpointing is implemented on multi versiong, thus space overhead is necessary. Since multi versioning is done on tuple level, and version sharing is used, the overhead is reduced. At present, the capacity of main memory is getting larger and larger, the cost is reasonable, because the efficiency of transaction processing and checkpointing is improved.A partition based parallel recovery algorithm is proposed to provide system availability during recovery. During recovery, recovery of individual sites is independent, parallelism of three types, namely parallelism between partition, parallelism between log disks, and parallelism between data and log, are exploited to speedup recovery, the total recovery time is cut down.The author has used the J-SIM software package to build a simulation system and conducted a seria of experiments, the feasibility and efficiency of the scheme proposed in this dissertation are verified. Experiment results show that: (1) Transaction response time is cut down owing to parallel log writing, from 50ms to 21 ms when log disk number is 8. (2) Total recovery time of the failure site is reduced from 65s to 28s. (3) The scheme achives higher performance than fuzzy checkpointing when the system is performing checkpointing. Query throughput of the scheme improves by about 67 percent over fuzzy checkpointing, and update transaction throughput improves by about 7.8 percent over fuzzy checkpointing. (4) The space overhead is around 11 percent in update intensive senarios, the overhead is acceptable. (5) The final experiment is conducted to measure system throughput and transaction response times during recovery.during four quarters of the recovery, system throughputs are 90.2Ktps,98.3Ktps,104.5Ktps,107.7Ktps, and avarage transaction response time are 273ms,32.3ms,9.2ms,5.32ms respectively.Experimnets results have verified the effetiveness and efficiency of the proposed scheme.The dissertation includes 49 digrams, and 5 tables, and refers to 121 papers. |