Research On Data Access Acceleration For NVM And RDMA

In recent years,with the development of science and technology,a number of novel hardware have been invented with greatly improved performance,such as Non-Volatile Memory(NVM)and Remote Direct Memory Access(RDMA)network interface cards.Compared with the traditional DRAM,NVM has a higher density and a larger capacity,but it also has limits such as fast reading and slow writing,a limited number of writing,etc.Compared with the traditional network transmission,the CPU and Kernel can be freed due to RDMA,but RDMA also brings other problems,such as memory management.This thesis revolves around the two new types of hardware,NVM and RDMA,to study the acceleration of data access based on NVM and RDMA.For the characteristics of NVM,this thesis proposes a new dynamic hash index to reduce the number of writing memory.It has excellent space utilization and also solves the consistency problem of data due to persistence.In response to the issue of RDMA on the management of the host server,this thesis proposes a new management strategy to minimize the number of network transmission times in one writing operation under the premise that the data is successfully read by the host.In general,the main work and contribution of this thesis can be summarized as follows:(1)A new NVM-based dynamic hash index NVM-LH and its optimization NVM-LH+is proposed.After considering the characteristics of NVM and the advantages and disadvantages of the traditional linear hash index,this thesis proposes a new NVM-friendly dynamic hash index NVM-LH.The index achieves the friendship of the cacheline through the address alignment and optimizes the number of memory reading and writing.The algorithm puts forward the problem of consistency for data persistence to ensure that consistency is suitable for linear hash indices.After discovering and analyzing the advantages and disadvantages of NVM-LH through evaluation,this thesis further optimized NVM-LH and proposed a new NVMfriendly dynamic hash index NVM-LH+.Under the premise of retaining the existing advantages of NVM-LH,the algorithm has designed a new bucket structure for NVM,which not only reduces additional reading and writing operations but also maintains high space utilization and less memory writing.NVM-LH+also remains the consistency guarantee methods from NVM-LH.This thesis conducts comparative experiments on a real NVM system to compare the performance of NVM-LH,NVM-LH+with the existing index CCEH(Cacheline-Conscious Extendible Hashing).The results show that both NVM-LH and NVM-LH+ outperform the competitor in memory writing and space utilization.(2)A new RDMA-based host memory management algorithm HMMM is proposed.After researching the characteristics of RDMA,this thesis proposes a new RDMAbased host memory management algorithm HMMM.It runs on a single thread to reduce CPU usage.After referring to the implicit list structure widely used in the computer system,we design the transmission strategy of front and rear coverage.In this case,the lazy update strategy was designed to reduce the number of network transmissions.This thesis conducts comparative experiments to compare the performance of HMMM with existing methods L5(Low-Level,Low-Latency Library)and FaRM(Fast Remote Memory).The results show that HMMM outperforms the competitors in both throughput and latency.