Optimization for QoS on Systems with Tasks Deadlines Luis Fernando Orleans Pedro Nuno Furtado
Introduction QoS in information systems is all about providing guarantees that submitted tasks will execute within an specified amount of time (deadline) A simple solution to reduce the response time is to add more servers Load-balancing Problem: no concerns about time- constraints
Introduction (2) Unpredictable behaviour of stressed parallel systems Possible causes: Load-balancing algorithms without concerns on response time (best-effort) Workload variability
Introduction (3) Objective: Optimize the performance of stressed QoS parallel systems. Methodology: Simulation
Some considerations Hard deadlines x soft deadlines Generation of tasks duration Exponential distribution Pareto’s distribution (a more real modelling) Tasks arrival rate: 10 tasks per sec. Exponential distribution
Load-balancing algorithms Round-Robin Requests are dispatched round-the-table among the servers Least-Work-Remaining Requests are dispatched to the server with the least outstanding work.
Algorithms with best-effort policy
Proposed solution Limit the number of concurrent executing (CE) tasks in the system Admission control Rejection of incoming tasks when the max CE had been reached
Results Treating the system as an M/M/3/CE Processor Sharing system
Results Treating the system as an M/P/3/CE Processor Sharing system, where P is a Pareto distribution
Conclusions In QoS system, the number of CE is a crucial variable Limiting the value of CE can drastically reduce the number of killed tasks Task size’s variability has an enormous impact on the system’s performance
Questions?
Thank you!