Research On High Performance Distributed Persistent Memory Systems

With the vigorous development of fields such as social media and artificial intelligence,the total amount of global data has grown exponentially.Data-intensive systems have gradually replaced computing-intensive systems as the mainstream of applications.These data-intensive systems place high demands on storage capacity and read/write performance.Therefore,it is necessary to build large-capacity memory systems and keep all the required data in memory.The current primary medium of main memory is difficult to further expand the capacity due to technological limitations.Moreover,the high copy overheads of the TCP/IP protocols limit the throughputs of memory and storage systems under the conditions where the network link speeds are high.In the current context,large-capacity and byte-addressable Persistent Memory(PM)and Remote Direct Memory Access(RDMA)network technologies that can bypass the system kernel and support zero memory copy have received extensive attentions.At the same time,building distributed RDMA-based Persistent Memory systems brings new development opportunities and challenges for the fast access and mass storage of big data.However,existing works fail to effectively address the problems in terms of the consistency guarantee for remote data access,the synergized read/write optimization for RDMA and PM,and the energy supply optimization for persistent memory systems.This dissertation investigates the above problems and proposes efficient solutions.Existing works introduce high system overheads when providing consistency guarantee for PM and RDMA-based remote data access.In order to address the problem,this dissertation proposes a PM-based log-structured design with consistency guarantee,called Erda.For the write requests,after obtaining the metadata,clients use the one-sided RDMA operations to directly write the data to the destination memory addresses in the persistent memory on servers without redundant copies.For the subsequent read requests,clients directly read the data via one-sided operations,and verify the integrity of the data via checksums without client-server coordination.Since the server retains the previous versions of data,and the metadata contains the addresses of these previous versions of data,once a client’s read procedure detects a failure,the system can efficiently restore to a consistent state.Experimental results show that compared with existing solutions,Erda significantly reduces CPU costs at server side,reduces the number of bytes written to PM by about 50% and the latencies by 30.6%～36.9%,and achieves 1.4X～1.6X throughput improvements.Existing index schemes separately optimize the PM write operations or the RDMA read operations,and lack synergized read-write optimization.In order to address the problems,this dissertation proposes a PM index structure called Continuity-Hashing to optimize both read and write operations.Continuity-Hashing designs a contiguous fine-grained shared region between each two adjacent hash buckets,which provides standby positions for the neighbouring buckets in case of hash collisions.Therefore,all the potential storage positions for a key-value pair are in a contiguous memory region,which can be fetched remotely with a single RDMA read operation.Continuity-Hashing assigns a data validity indicator that is less than the size of an atomic write unit for each two adjacent hash buckets and the shared region between the two buckets,and is able to provide log-free consistency guarantee for all the PM write operations.In addition,Continuity-Hashing also proposes an optimization scheme for the space utilization.Experimental results show that compared with existing schemes,Continuity-Hashing achieves the smallest number of PM writes,1.73 X speedup on average and 1.2X～3.1X throughput improvements for various workloads.To cope with the high energy consumption of PM reads and writes in the era of big data,from the perspective of efficient energy supply,this dissertation proposes a dynamic scheduling method for stable energy supply in PM systems,called Smoother.Smoother identifies the energy supply as a stable supply mode and a fluctuating supply mode.With a low-power energy storage device as a stabilizer,Smoother computes a nonlinear programming problem with linear constraints to perform real-time stabilization for the energy in the fluctuating supply mode.In addition,Smoother balances the energy supply and the energy demand of the workloads by dynamically scheduling PM-based workloads according to the stable energy supply.Experimental results show that compared with existing solutions,Smoother significantly reduces the negative impact of energy fluctuations on the stability of PM systems,improves the utilization of clean energy by an average of 1.81 times,and decreases the electricity costs by an average of 54.46%,thus providing a stable and reliable energy supply for PM systems.