Aggregate Scheduling Enhancing Throughput in Collective Tasking Systems

Aggregate Scheduling – Enhancing Throughput in Collective Tasking Systems L. Subramanian Randy H. Katz Michael J. Franklin

Collective Tasking Systems l Properties : – – l Services requests of a predefined set of types Every request has an associated type All requests of a particular type can be aggregated into a single request Bottleneck operation of every type is performed only once for all requests of that type Examples: – – – – Broadcast disks – application of broadcast scheduling. Reservation systems – access to the reservation database Network Provisioning systems – bandwidth brokers Front-end Database monitors –access point for multiple databases Disk scheduling systems –locality based access in disks Caching Systems Gang Scheduling – Multiprocessor systems

Aggregate Scheduling Scheduler application List of Queues bottleneck OPT Door Maintainer Aggregator List of Queues: A queue of requests for every type OPT: Aggregate Statistics of requests of every type Doorkeeper: Triggers event when a new request arrives

Components in an Aggregate Scheduling System Aggregator: • Aggregates requests into types • Updates OPT data structure • Informs Maintainer about new event Scheduler: • Computes the type with maximum value of OPT function • Computes Aggregate request for all requests of that type • Schedules that type to the application Maintainer: • Uses an optimization function for types • Maintains the invariant property of OPT for new events OPT: • Data Structure optimized for the optimization metric • Every optimization metric induces an invariant in OPT

Optimization Metrics Rx. W scheduling – (#of Requests) * (Max Waiting Time) l Approximate Rx. W – Apply Rx. W for reduced set of types l Kinetic Tournaments – Total waiting time for requests in a queue l Gang Scheduling l – Associate distance metric between processes (frequency of IPC) – Schedule group of processes with min value of max distance The Cost Dimension – Cost associated with every type (cost of bottleneck operation) – Costs can be dynamic (eg. disk scheduling) – Fagin’s work on fuzzy systems l Other variants l – Bounded queue size (admission control) – Bounded response time (earliest deadline)

Network Provisioning System • 12 basic domains in AT&T’s backbone • 10% of bandwidth reserved(statistically) for Vo. IP and VPNs. • A provisioning system accepts interdomain requests and reserves along a path. • All requests between a pair of domains are aggregated into a single request. • Regulate traffic for the reserved portion.

Throughput & Block Rate Characteristics

Response Time Characteristics

Conclusions Rx. W and Kinetic tournaments give much better performance than FIFO l Rx. W vs Kinetic Tournaments(KT) l – – – l Rx. W has slightly higher throughput than KT KT has much lesser response time at operating range Variation of response time in KT is restricted Max response time of KT is very low (6 times) Rx. W has starvation problem Experiment aggregate scheduling for other collective tasking systems