Fast (re-)configuration of mixed on-demand and spot instance pools for high-throughput computing

Abstract

Commercial cloud offerings let users allocate compute resources on demand, charging based on reserved time intervals. Users, however, lack guidance for assembling instance pools from different cloud instance types, in order to control completion time and monetary budget. BaTS, our budgetconstrained scheduler uses tiny statistical samples of task executions in order to predict completion times (and associated costs) for given bags of tasks, allowing the user to favor either fast execution or low computation budget. BaTS' estimator, however, can not handle variably-priced spot instances appropriately.In this work, we present a new prediction module for BaTS that quickly computes accurate approximations to the Pareto set of mixed on-demand and spot instance pools, based on a genetic algorithm (GA). This new approach allows BaTS to react to changing spot instance prices at runtime by re-configuring the instance pool according to the user's runtime and budget constraints. Simulator-based results show that the GA can approximate the Pareto sets for machine configurations in about 30 seconds time, without noticeable loss of quality, compared to an exact solution, computed offline within 40 to 60 minutes time.

Fast Pareto Front Approximation for Cloud Instance Pool Optimization

“…We compare the MOO-GA fast, well-approximated Pareto sets (PS) of cloud schedules to the exact PS [3] and the NS-GAIII [13] approximations. Chosen PS metrics show domainspecific desired qualities and computational metrics, timeliness.…”

“…Chosen PS metrics show domainspecific desired qualities and computational metrics, timeliness. Test problems are domain-specific workloads [3]. Workloads: An on-demand instance type (OD) has an hourly price and execution speed; its related spot type (S) [2] at $0.020; 2) 2x faster at $0.065; 3) 3x faster at $0.130; the related S types cost $0.003, $0.007 and $0.013.…”

“…In this work, we introduced heuristics derived from cloud scheduling to help MOO-GA, our extended genetic algorithm [3] used by BaTS [2], deliver fast, well-approximated Pareto sets. Results show that exact Pareto set computation is unfeasible for online instance pool re-configuration, while our heuristics-enhanced MOO-GA computes controllable-quality approximations in less than one second time and of better quality than state-of-the-art NSGAIII.…”

“…However, allocating the right number of instances, of the right type, for the right time, strongly depends on the application, and is left to the user. State-of-the-art cloud scheduler BaTS [2] and its fast GA approximate within seconds Pareto sets of makespan-cost options [3]. However, this GA lacks an adaptive termination criterion enforcing solution quality.…”

E-BaTS: Energy-Aware Scheduling for Bag-of-Task Applications in HPC Clusters

“…BaTS does not need apriori knowledge about the bag; it only needs to know the total number of tasks in the bag. To approximate the best combinations of machines presented to the user, in terms of makespan and cost, which form a Pareto Set, BaTS uses a genetic algorithm (GA) [27].…”