Design And Implementation Of Distributed Cache For Heterogeneous Multilevel Storage

Distributed caching systems have become an indispensable infrastructure in the Internet industry due to their high concurrency and low access latency.However,because caching systems are mostly built based on expensive memory,they are expensive to deploy and operate and maintain.With the development of hybrid storage,the use of DRAM and SSD-based multi-level hybrid storage architecture provides a new development direction for caching systems,but further exploration is needed on the strategies for data placement,migration and backup.In this thesis,we design and develop a distributed caching system based on heterogeneous multilevel storage to address the shortcomings of the above research,which realizes hot and cold data tiering,automatic migration and persistence of in-memory data,and further optimizes the distributed caching system for disadvantages such as low memory utilization,difficult operation and maintenance,and inefficient indexing.The main work of this thesis are as follows.(1)A tiered storage strategy for hot and cold cache based on heat statistics is proposed for hardware environments that employ DRAM and SSD.A hybrid storage architecture with cold data stored in SSD and metadata and hot data stored in DRAM is realized through the automatic hot and cold data migration strategy.The data persistence module based on SSD storage module is designed and implemented based on LSM-Tree with key-value separation,which ensures efficient persistence while reducing the additional wear and tear overhead on flash particles from LSM-Tree compaction.(2)A hash index structure based on cuckoo hash is designed and implemented to address the inefficiency of zipper method hash table index access.Using the 64-bit processor in-core cache prefetch mechanism and virtual memory effective pointer bits,the hash bucket data structure is optimized to improve memory utilization while effectively reducing the number of memory accesses when performing hash table access.To further improve memory utilization,this thesis proposes a memory allocation and reallocation strategy based on memory usage load statistics and cuckoo filters to improve the cache hit rate of the system by allocating more memory to hot and high-load task modules.(3)To address the problem that existing open source caching systems are difficult to operate and maintain,a centralized distributed caching architecture is designed and implemented to collect and manage the state of cache nodes in the system through metadata clusters,thus providing real-time cluster status to administrators as well as clients.Finally,the distributed caching system for multi-level heterogeneous storage is tested for functionality and stress.The analysis of the test results shows that the system in this thesis has the necessary high concurrency and high availability requirements for caching systems and meets the design objectives of the system.