Cloud operators utilize collective communication optimizers to enhance the efficiency of the single-tenant, centrally managed training clusters they manage. However, current optimizers struggle to scale for such use cases and often compromise solution quality for scalability. Our solution, TE-CCL, adopts a traffic-engineering-based approach to collective communication. Compared to a state-of-the-art optimizer, TACCL, TE-CCL produced schedules with 2× better performance on topologies TACCL supports (and its solver took a similar amount of time as TACCL's heuristic-based approach). TECCL additionally scales to larger topologies than TACCL. On our GPU testbed, TE-CCL outperformed TACCL by 2.14× and RCCL by 3.18× in terms of algorithm bandwidth.

Many resource management problems require sequential decision-making under uncertainty, where the only uncertainty affecting the decision outcomes are exogenous variables outside the control of the decision-maker. We model these problems as Exo-MDPs (Markov Decision Processes with Exogenous Inputs) and design a class of data-efficient algorithms for them termed Hindsight Learning (HL). Our HL algorithms achieve data efficiency by leveraging a key insight: having samples of the exogenous variables, past decisions can be revisited in hindsight to infer counterfactual consequences that can accelerate policy improvements. We compare HL against classic baselines in the multi-secretary and airline revenue management problems. We also scale our algorithms to a business-critical cloud resource management problem – allocating Virtual Machines (VMs) to physical machines, and simulate their performance with real datasets from a large public cloud provider. We find that HL algorithms outperform domain-specific heuristics, as well as state-of-the-art reinforcement learning methods.

In Database-as-a-Service (DBaaS) clusters, resource management is a complex optimization problem that assigns tenants to nodes, subject to various constraints and objectives. Tenants share resources within a node, however, their resource demands can change over time and exhibit high variance. As tenants may accumulate large state, moving them to a different node becomes disruptive, making intelligent placement decisions crucial to avoid service disruption. Placement decisions need to account for dynamic changes in tenant resource demands, different causes of service disruption, and various placement constraints, giving rise to a complex search space. In this paper, we show how to bring combinatorial solvers to bear on this problem, formulating the objective of minimizing service disruption as an optimization problem amenable to fast solutions. We implemented our approach in the Service Fabric cluster manager codebase. Experiments show significant reductions in constraint violations and tenant moves, compared to the previous state-of-the-art, including the unmodified Service Fabric cluster manager, as well as recent research on DBaaS tenant placement.

Large food-service companies typically support a wide range of operations (catering, vending machines, repairs), each with different operational characteristics (manpower, vehicles, tools, timing constraints, etc.). While the advances in Internet-based technologies facilitate the adoption of automated scheduling systems, the complexity and heterogeneity of the different operations hinders the design of comprehensive optimization solutions. Indeed, our collaboration with Compass Group, one of the largest food-service companies in the world, reveals that many of its workforce assignments are done manually due to the lack of scheduling solutions that can accommodate the complexity of operational constraints. Further, the diversity in the nature of operations prevents collaboration and sharing of resources among various services such as catering and beverage distribution, leading to an inflated fleet size.

In this paper, we describe Multi-Itinerary Optimization (MIO) - a novel Bing maps service that automates the process of building itineraries for multiple agents while optimizing their routes to save travel time or distance. MIO can be used by organizations with a fleet of vehicles and drivers, mobile salesforce, or a team of personnel in the field in order to maximize workforce efficiency. MIO accounts for service time windows, duration, and priority, as well as traffic conditions between locations, resulting in challenging algorithmic problems at multiple levels (e.g., calculating travel-time distance matrices at scale, scheduling services for multiple agents).