Research On Hot-Data-Aware Cache Strategy In Serverless Computing

In serverless computing,applications are split into a series of cloud functions,which run in different containers.As a result,serverless computing achieves on-demand resource allocation and fine-grained auto-scaling.However,this also leads to complete decoupling of computing and storage,so cloud functions need to share data through remote storage.Stateful applications are divided into multiple execution stages.Due to data sharing among execution stages and lack of system resources,data generally has long read intervals.Traditional cache strategies evict these data incorrectly,making it impossible for remote storage based on such strategies to achieve both high performance and low cost.According to experiments,traditional cache strategies have the following drawbacks:firstly,stateful applications have a large amount of once-read data and traditional cache strategies directly cache these data,resulting in the eviction of data with long read intervals,which reduces cache hit rates;secondly,multi-read data is shared by concurrent cloud functions,resulting in multi-read data being read frequently and then quickly becoming cold,which cannot be evicted immediately by traditional cache strategies,reducing memory utilisation;thirdly,there is correlation amang multi-read data,which is read periodically.However,traditional cache strategies cannot effectively pre-fetch these data.To solve the above problems,MListCache,a cache replacement strategy based on data hotness,is proposed,the design principle of which is to cache the multi-read data as soon as possible,while reducing the eviction of data with long read intervals.Firstly,MListCache records the read time of different data to bypass once-read data,and loads the multi-read data directly into memory when it appears for the first time.Secondly,to quickly evict data becoming cold,MListCache predicts the different data’s probability of next read based on the change in the data read intervals,and preferentially evicts data with a lower access probability to reduce evictions of data with long read intervals.Finally,MListCache establishes the correlation between different data.When a read request for a certain data comes,MListCache will pre-fetch all its correlated data.Compared with traditional cache strategies,MListCache achieves up to 2.74 x hit rate,reduces remote read latency by up to 80%,and reduces long-tail latency by 40%.Compared to native Min IO,MListCache reduces remote data read latency by 90%,while introducing only 13% external cost.Compared with Redis,MListCache achieves 90% of its throughput while saving 71% cost.