Our partners

CNRS

Search





Home page > English > Departments > Data Management > PYRAMID Team > Research

Performance Isolation

 

From the DaaS (Database-as-a-Service) provider’s point of view, multi-tenancy allows sharing resources of a single database server, in order to achieve the cost-effective goal. However, for an individual tenant, the performance can be degraded because of the interference of other tenants. Resource management techniques developed for the traditional DBMS are mostly based on relative priorities or proportional sharing. These techniques are inadequate in a public cloud setting: the amount of resources dedicated to a tenant depends on the activities of other co-located tenants; hence, there is no way to guarantee a service level agreement (SLA) in terms of performance. A possible solution is to enforce performance isolation, i.e. to specify the performance SLA by fixing the absolute amount of required resources (e.g. CPU, I/O, memory, and network) and the penalty in case of violation. Due to the changing workload, dynamic resource reservation mechanisms are needed. Other technical challenges of performance isolation include: (1) how to meter violation of performance SLAs for each type of resources with low overheads; (2) in case of overbooking, how to make the trade-off between reducing penalties and fairness to tenants; (3) how to deal with the discrepancy of the tenant required resource amount and the query optimizer estimated resource consumption .