Scalable Caching System For Short Videos

With the development of short video applications,the scale of short videos is growing rapidly,so the caching systems in Content Delivery Network(CDN)for short videos face challenges.These short videos often exhibit skewed data accesses,frequent insertion/deletion,which limits the system performance.To ensure rapid responses to hot video requests,the proxy servers in traditional CDN require big capacity memory for storing hot objects to obtain high Object Hit Ratio(OHR)for hot object caches.Existing proxy servers contain high-cost DRAM as hot object cache(i.e.,memory),but suffer from cachelevel write amplification due to the limited capacity for DRAM.Non-volatile Memory(NVM)has comparable access latency and lower cost than DRAM,while offering larger capacity,so NVM can be an ideal choice for proxy servers' memory.However,NVM has limited write endurance that is exacerbated by the frequent use of short video files,thus shortening the lifetime of NVM.To address this problem,a Scalable Caching System for Short Videos(SpeCache)is proposed.The hot object cache of the proxy server is composed of DRAM and NVM hybrid memory to guarantee high OHR for hot short videos.The proxy servers are grouped and share data in groups to reduce the number of writes to the NVM.To avoid any proxy server to become the bottleneck within a group,SpeCache leverages traffic-storage load balance scheme to offload hot videos from high-traffic proxies to light-traffic proxies by exploiting request statistics and possible vote.Moreover,SpeCache uses Bloom filter(BF)replicas to efficiently reduce access latency for users.Our design avoids the inconsistency between BF replicas and their corresponding data on proxy servers by differentiating the reason of BF replicas updates and rank these updates to ensure consistency.Experimental results demonstrate that compared with state-of-the-art CDN caching system,SpeCache efficiently reduces the number of writes to NVM by 25%,and reduces the access delay for users.