An Adaptive Hot-Cold Aware Two-Level Cache Mapping Framework for Distributed Caching Systems

At present, most of the existing distributed cache systems use consistent hash to map data to nodes. However, the existing consistent hashing schemes focus on the data migration overhead in the process of node change, and ignore the access frequency skew problem in the real workload. When a small number of hot keys carry a large number of requests, the conventional hash mapping will lead to uneven node load, waste of resources, and seriously affect the performance of the cache system. To solve the above problems, this paper designs an adaptive cold and hot aware double-layer cache mapping framework. This framework first divides the data into cold data and hot data according to the access heat of the key, and then uses different mapping strategies for these two types of data. For cold data, stable mapping is used to improve cache affinity and reduce migration overhead. For hot data, the access pressure is dispersed through the dynamic allocation path at the request level. At the same time, an adaptive adjustment mechanism is added at runtime to dynamically complete the upgrade or downgrade of hotspots; From the experimental results, in the scenario of high hot spot tilt workload, this scheme can reduce the node load balancing coefficient to 3.47%, and also maintain stable performance when the dynamic load changes. The ablation experiment further shows that there is a synergy between the hot and cold double-layer mapping structure and the dynamic load sensing mechanism.