SLO Oriented Dynamic Balance Mechanism Of Resource In Multi-tenant Environment

The demand for SaaS(Software as a Service)multi-tenant applications in the cloud computing era is growing.Multi-tenant data management is the key to rapid development and efficient operation of SaaS applications.Multi-tenant data management saves the cost of data management and maintenance through shared resources.But from the perspective of tenants,each tenant will negotiate with a Service Level Agreement(Service Level Agreement,SLA)with the suppliers.In this paper,we focus on the performance metrics associated with the SLA-the response time of the query,so multi-tenant data queries still need to meet tenant's SLO(Service Level Objectives),which describes the set of benchmarks or targets set by the parties,and relates to the service provided by the supplier for a given period of time.The number of suppliers who use this approach to provide services to tenants gradually increases.The multi-tenant data management consolidate tenants on a shared database is an efficient way of reducing resources and costs.But the resource sharing between tenants also brought a series of problems and challenges.First of all,in order to improve the availability and fault tolerance of the system,the tenant have multiple replicas.The tenant's replicas place at different node are not same.Different tenant have different resources demand because of their different nature of business.Some tenants have more demand for CPU resources,less demand for I/o resources,and some tenants are just the opposite.Unreasonable query scheduling can lead to an imbalanced usage of resources among different servers,and it could cause resource waste and significantly degrade the overall performance.For instance,co-locating multiple CPU-heavy queries is probably worse than co-locating a mix of I/O-heavy and CPU-heavy queries.Secondly,the multi-tenant data access load has the characteristics of mixed,large fluctuation and multi-change and so on.Resource sharing in multi-tenant databases can adversely affect a tenant's performance due to contending for shared resources among other tenants.A node with many tenants or tenants with heavy workload will affect the performance of the tenants on the node.When a tenant receives an increased workload,the other tenants on the same node will suffer from the increased total workload,and cannot meet their SLOs.The increased total workload will make the node to be overloaded,and may cause a performance crisis.Therefore,how to balance resource utilization across nodes and eliminate the performance crisis is becoming a growing concern for suppliers.This paper is based on the actual needs of users,conducted a series of studies on the overload problem in the multi-tenant environment for the existing problems of the existing work,and proposes a dynamic balance mechanism of resource to balance resource utilization across nodes and eliminate the performance crisis.The specific work and contributions of this paper are summarized as follows:1.Proposed three load models and describe the construction process of the three load models in detail.This paper defined three load models:query load model,tenant replica load model and node load model which expressed the load of query,tenant and model respectively.First,we use supervised learning techniques to learn characteristics of queries and train query load model.The load of a query is defined as the amount of resources required to serve the query,as a percentage of the server capacity.Then we use a linear additive method to train tenant replica load model and node load model.Finally,the accuracy of the load model is verified by experiments.2.For the problem of the unbalanced resource usage among different nodes which caused by the unreasonable query scheduling strategy.This paper presented a Dynamic query scheduling strategy which based on the load model and the nodes performance.First,in order to monitor the performance of nodes in real time,in this paper,server performance labels are trained corresponding to the load level of servers.5.The dynamic query scheduling policy is then used to detect and count the performance labels of all nodes in real time.Dynamic scheduling the tenant's query based on the query load model and node performance labels,choosing a appropriate node which contains the tenant's replica to execute the query.This strategy can balance resource utilization across nodes and improve node resource utilization.3.For the problem of performance crisis in the multi-tenant environment due to tenant's increased workload.This paper presented a lightweight load balancing mechanism to eliminate node performance crisis.In a multi-tenant database environment,a performance crisis could make the tenant's resource needs cannot be met and see violations on their SLO.In this paper,a lightweight load balancing mechanism is presented to address the performance crisis problems in light of exchanging the roles between tenants' primary replicas and secondary replicas.The principle of eliminating the performance crisis is as follows.Because each tenant has one primary replica with one or more secondary replicas.Both the primary and secondary replicas in the mechanism can serve "read-only" queries,and meanwhile,the primary replica can serve"write" queries as well.Hence,the primary replica receives a larger amount of workload compared with the secondary replica.After tenants perform the replica exchange,workload can be effectively moved from the primary replica to the secondary replica,and the performance crisis is mitigated.Finally,the roles of the primary and the secondary replicas are exchanged and the "write" queries are sent to the new primary replica,which was a secondary replica before the exchange.The effectiveness of the dynamic balance mechanism of resource is verified by experiments,the mechanism can balance resource utilization across nodes and eliminate the performance crisis quickly and effectively.