How Alibaba’s Sigma‑Cerebro Simulator Boosts Cluster Utilization for Double‑11

Alibaba’s Sigma is a Pouch‑based container scheduling system used across the entire Alibaba group. Since its 2017 launch, Sigma has supported all containers during Double‑11, cutting IT costs by 50% and becoming a critical infrastructure for Alibaba’s operations.

Sigma‑Cerebro is a simulation platform that mimics 1:1 production machine resources and request demands without real hosts, enabling rapid, low‑cost evaluation of scaling algorithms under resource fragmentation, large‑scale scaling, and unexpected over‑sell scenarios.

During Double‑11 2017, Cerebro pre‑processing allowed Sigma to complete a one‑click site build in 30 minutes, raising static allocation from 66% to 95% and greatly improving resource efficiency.

What makes a good scheduler?

A good scheduler minimizes interference while maximizing overall cluster resource utilization and completing allocations within fixed time windows.

Requirements for a scheduling simulation system

Accurately simulate large‑scale production environments and complex demands.

Minimize simulation overhead and risk.

Provide both static and dynamic quantitative answers.

The simulator uses background data stored in OSS, requiring only data retrieval during simulation. Multiple environment pools allow a full task set with just three containers per pool.

Supported modes include scaling algorithm evaluation, pre‑allocation, and issue reproduction.

Users can configure custom watermarks and schedulers; the simulator injects identical 1:1 host data and request workloads, runs the user’s algorithm, and scores the results, enabling comparison of different algorithm versions.

Why a scheduling simulator?

Key challenges in container scheduling:

Measuring the quality of scheduling decisions.

Overcoming slow image pulls during massive one‑click site builds.

Accurately estimating resources for large‑scale simultaneous builds.

Reproducing production scheduling issues in test environments.

The simulator addresses these by providing realistic background data and request workloads for clear static resource allocation assessment.

Measuring scheduling quality

Example: two hosts each with 4 free CPU cores. Three containers (A:2 CPU, B:2 CPU, C:4 CPU) illustrate how a poor algorithm can leave C unscheduled, achieving only 50% static allocation, whereas an optimal algorithm achieves 100%.

When multiple resources (CPU, memory, disk) are considered, fragmentation can cause hosts to run out of one resource while others remain idle, leading to waste.

Additional complexities include affinity, anti‑affinity, exclusive requirements, and disaster‑recovery constraints, further complicating the allocation problem.

Handling rapid scaling with limited host I/O

Image download and extraction dominate container startup time, often exceeding 50% of total latency. Alibaba’s Pouch uses a P2P “Dragonfly” system for efficient image distribution and supports pre‑loading to reduce startup time.

However, limited host disk capacity can prevent pre‑loading all large images, and weak disk I/O may cause timeouts even with network‑optimized distribution.

Precise pre‑allocation via the simulator can identify which containers need pre‑loading, mitigating these issues.

Resource demand forecasting

Fragmentation can inflate the perceived need for additional hosts; accurate simulation helps estimate the optimal number of hosts to avoid over‑provisioning.

Reproducing production scheduling issues in test environments

The simulator enables faithful recreation of production scenarios, providing clear guidance for issue resolution without impacting live services.

Future plans

Current simulations are static; future work includes dynamic orthogonal simulations to complement static analysis, exploring optimal microservice combinations, and leveraging cpushare for better peak‑shaving.

Ultimately, mixed‑workload elastic capabilities will be opened to Alibaba Cloud users, delivering stronger compute power at lower cost and improving overall resource efficiency.