How Scheduling Algorithms Power Efficient Data Center Resource Management

Resource management systems abstract data center resources and must ensure application stability, performance (SLA), efficiency, and energy savings. Scheduling algorithms are a key component that decides on which machine a compute task should run.

Internet Applications and Modern Data Centers

Cloud computing powers many services; large cloud providers operate many data centers with numerous physical servers. To manage these servers, a Cluster Resource Management System (CRMS) is needed, whose value can be described as "Datacenter as a Computer".

Value of Scheduling Algorithms

Scheduling algorithms determine the placement of tasks in a cluster.

In containerized environments, the scheduler places container instances (e.g., Docker, PouchContainer) onto suitable hosts, providing benefits at three levels.

Container‑level Benefits

Meet resource requirements (CPU, memory, disk, network, special OS or hardware).

Provide a comfortable environment by avoiding resource contention between containers.

Application‑level Benefits

High availability: multiple instances run simultaneously so a single failure does not impact the service.

Disaster tolerance: instances are spread across hosts, racks, rooms, data centers, cities, and even countries.

Advanced placement requirements such as ordering, data locality, etc.

Data‑center‑level Benefits

Cost reduction: efficient packing reduces the number of servers needed, lowering hardware, space, power, and cooling expenses.

Additional considerations include fairness, inter‑application interference, resource sharing, and single‑machine allocation (e.g., hyper‑threading, memory bandwidth). Alibaba’s production system Sigma uses complex scheduling rules.

Alibaba Global Scheduling Algorithm Challenge

The competition presents a simplified real‑world scenario with about 6 000 hosts and 68 000 instances. Constraints include resource limits, high‑availability groups (P, M, PM), and anti‑affinity between applications.

Resource Constraints

Each instance has CPU and memory requirements that vary over a 24‑hour curve, creating optimization opportunities and complexity.

High‑Availability Constraints

Important applications are labeled P, M, or PM, and limits on how many such instances may share a host ensure minimal impact from host failures.

Anti‑Affinity Constraints

Pairs of applications have a limit k on how many instances of the second can co‑locate with an instance of the first, reducing performance interference.

Optimization Objective

The goal is to keep per‑host resource utilization within a target range while minimizing the number of active hosts, thereby saving cost and preserving headroom for load spikes.

Invitation

Researchers and engineers interested in resource scheduling, optimization, and algorithms are invited to participate for prizes and a chance to attend a hackathon in the United States.