A Heterogeneous Access-aware Scheduling Scheme For In-memory Stores

In recent years,to improve the user experience of latency-sensitive applications,more and more data is stored in the memory of the cluster system,in-memory stores are derived.In-memory stores are affected by various heterogeneous access factors,and existing scheduling schemes usually only consider a single heterogeneity factor,so they cannot solve the performance issue caused by heterogeneous accesses of in-memory stores very well.To address the above problem,this paper investigates a scheduling scheme for in-memory stores.After in-depth analysis,heterogeneous access factors affecting in-memory stores include at least data location(local or remote rack),node load(light or heavy load)and block state(cache hit or miss).With taking the above three heterogeneous access factors into account,the paper proposes a unified assessment metric of ‘access cost' to quantitatively evaluate heterogeneous access overhead.Since in-memory stores may have different optimization goals,the paper proposes three optimization scheduling strategies(Minimal-Latency,Minimal-Traffic and Maximal-Balancing)based on access cost.As for an access request,the basic idea of Minimal-Latency is to select the one with the least cost of normal read,disk read and degraded read;the basic idea of Minimal-Traffic is to select the one with the least cost of normal read and disk read;Maximal-Balancing is a variant of Minimal-Latency,where the access cost of disk read ignores the time overhead of accessing the disk.On the basis of the above optimization scheduling strategies,the paper proposes a heterogeneity-aware scheduling scheme-HaSS-for in-memory stores.HaSS supports two-level scheduling optimization,the block-level scheduling is specific to each access request and can support each optimization scheduling strategy;the request-level scheduling is specific to a group of access requests and can reorder requests based on access cost.In addition,HaSS supports both in-memory stores adopting a hybrid redundancy scheme and in-memory stores adopting a single redundancy scheme.To comparatively evaluate the above three optimization scheduling strategies,a cache-hit-first strategy called Basic is also deployed by the HaSS scheduler.Quantitative experiments show that,under a heterogeneous-access scenario,the above three optimization scheduling strategies can accomplish the best performance in terms of access latency,network traffic and load balancing,respectively.For example,the verification experiments show that,Minimal-Latency can reduce the average access latency by 22% compared with Basic;the network traffic of Minimal-Traffic and Basic is the least;the RSS values of Maximal-Balancing and Basic are 0.06 and 5.20 respectively,where the smaller the RSS value,the better the load balancing.