LraSched: Admitting More Long-Running Applications via Auto-Estimating Container Size and Affinity

Abstract

Many long-running applications (LRAs) are increasingly using containerization in shared production clusters. To achieve high resource efficiency and LRA performance, one of the key decisions made by existing cluster schedulers is the placement of LRA containers within a cluster. However, they fail to account for estimating the size and affinity of LRA containers before executing placement. We present LraSched, a cluster scheduler that places LRA containers onto machines based on their sizes and affinities while providing consistently high performance. LraSched introduces an automated method that leverages historical data and collects new information to estimate container size and affinity for an LRA. Specifically, it uses an online machine learning method to map a new incoming LRA to previous workloads from which we can transfer experience and recommends the amount of resources (size) and the degree of collocation (affinity) for the containers of the new incoming LRA. By means of recommendations, LraSched adapts the heuristic for vector bin packing to LRA scheduling and places LRA containers in a manner that both maximizes the number of LRAs deployed and minimizes the resource fragmentation, but without affecting LRA performance. Testbed and simulation experiments show that LraSched can improve the resource utilization by up to 6.2% while meeting performance constraints for LRAs.

Affinity-Aware Resource Provisioning for Long-Running Applications in Shared Clusters

“…Finally, the Extended Sum is an adaptation of the measure used in LraSched [11]. For application i, it is defined as the sum, over all dimensions h, of the demands of all replicas of that application |R i |s ih in dimension h normalized by the total demand W h of all applications in that dimension.…”

“…Baseline Methods. We mainly have three baselines in the experiments: two heuristics of Medea [21] and one heuristic of LraSched [11].…”

“…LraSched [11] uses a two-phase approach. The first phase aims at maximizing the number of fully allocated LRAs and resource utilization of the given restricted pool of available nodes.…”

“…Kubernetes [1] and Medea [21] address the application-related affinity, but do not address the requirement of scheduling all LRAs as a global optimization problem: Kubernetes schedules one LRA replica (pod) at a decision point, while Medea aims at runtime scheduling of relatively small batches of LRAs periodically. LraSched [11] only addresses the intra-application affinity constraints. Additionally, the capability of handling massivescale scheduling problems of these three solutions has not been fully investigated.…”

“…This motivates us to develop an algorithm suite that can be used by practitioners for selecting the best performing heuristics that best fit the specific needs of a given scheduling scenario. To illustrate the capabilities of the suite, we perform experiments on instances generated from the Alibaba Tianchi dataset [44] and compare the winning approaches from our suite with the best performing published algorithms: two heuristics from Medea [21], namely TagPopularity and NodeCandidate, as well as a heuristic based on the Fitness measure introduced in LraSched [11]. A highlevel summary of the most successful algorithms in our toolkit and those published in the literature is presented in Fig.…”