Research On Parallel Integrity Tree Crash Recovery Strategy For Persistent Memory

The burgeoning non-volatile memory(Non-Volatile Memory,NVM)technology can meet the demand of massive memory capacity for data-intensive applications.However,NVM is more likely to cause data security problems due to its non-volatile nature.Generally,encryption and integrity verification are used to ensure system security.Parallel integrity tree can be verified and updated in parallel which is widely studied and applied accordingly.However,parallel integrity tree is more challenging to recover after crash.Persisting updates to secure metadata can reduce the crash recovery time.Nevertheless,persisting secure metadata will degrade system performance,due to the unpleasant write operation of NVM.Therefore,how to ensure that parallel integrity tree can be quickly recovered after a crash while not having an obvious impact on system performance is still a research focus.According to the different focus of fast recovery and high performance,this paper proposes two parallel integrity tree crash recovery schemes,called Hades and Hades-Perf respectively.Hades customizes different crash recovery strategies for different levels according to the write behavior of different levels of the parallel integrity tree.Anubis strategy is used for the upper nodes that are not frequently updated while the Osiris strategy is used for the lower nodes that are frequently updated.The Hades scheme can combine the advantages of little additional write requests caused by the Osiris solution and the short recovery time of the Anubis solution.In order to quickly locate the location of the nodes that need to be restored after crash,Hades adopts a multi-level bitmap index method to further reduce recovery time.In the Hades-Perf scheme,the Rhea and Osiris plus strategy are used for the odd and even nodes of the integrity tree respectively.Osiris plus improves write requests caused by metadata cache by reducing the evict of dirty even blocks,thereby further improving system performance.Odd nodes adopt the Rhea strategy which takes advantage of the extra resources provided by the ECC chip and vacant locations in metadata cache to assist neighboring layer recovery,while adding enhanced verification mechanisms to ensure system security.The two strategies,Hades and Hades-Perf,are implemented and tested in the simulation system built by Gem5 and NVMain.The results show that the IPC of Hades and Hades-Perf reach 1.17 times and 1.23 times that of the state-of-art Anubis scheme respectively.At the same time,the crash recovery time of Hades-Perf scheme is close to that of the Anubis scheme,while Hades can speed up recovery time by almost 2×10~5 times(17ms/75ns).According to the evaluation results,it can be concluded that the Hades and Hades-Perf scheme fit better in system which focus on high performance and high availability respectively.