RH-LSM:A Read-sensitive And Heterogeneous LSM-tree

As a write-optimized storage structure,Log Structured Merge Trees(LSM-tree)has been widely used in the design of modern storage engines.LSM-tree buffers write operations as log in memory and delays flushing to the disk as tiered storage.LSM-tree maintains a background task called compaction to shorten the read path and reclaim the storage space.The tiered storage of LSM-tree ignores the read pattern of the given workload.For those frequently accessed but barely updated data,it is placed at the bottom layer via compaction,which will bring in more disk IO and affect the read performance.Moreover,the procedure of compaction involves the decoding,encoding,comparison and merge operations.Essentially,compaction is a computation-intensive operation,which is prone to resulting in the performance thrashing problem when it is scheduled and competes for the CPU resource.In this paper,we firstly review the recent research on LSM-tree.To address the high read latency and performance thrashing problems of LSM-tree,we propose an enhanced storage structure RH-LSM.RH-LSM determines which cold data should be compacted according to the calculated access frequency via capturing the data historical read patterns.By this means we ensure that frequently accessed data has a shorter read path.Moreover,in RH-LSM,compaction is offloaded to the FPGA.Compaction offload includes efficient FPGA pipeline design and asynchronous compaction scheduler.To avoid the FPGA intrinsic flaws,a fault-tolerance mechanism is implemented to ensure the data correctness.RH-LSM is a storage structure that considers access patterns and has stable performance.The experiments on public benchmarks show that RH-LSM surpasses other state-of-art LSM-tree based storage engines.