How Dynamic Resource Scheduling Boosts OpenStack Efficiency and Cuts Costs

Background

Virtualization technology packages compute, storage, and network resources into isolated virtual environments, enabling sharing, isolation, and dynamic management. As virtualization adoption grows, efficient scheduling and management of these resources become critical, leading to the development of virtualization resource scheduling algorithms that aim to optimize system performance and resource utilization.

OpenStack NUMA‑Based Instance Model

In 360's OpenStack cloud solution, dedicated instances use a core‑binding mode to ensure CPU isolation and performance. Two modes exist: single‑NUMA (all vCPUs on one NUMA node) and dual‑NUMA (vCPUs split across two NUMA nodes). Typically, single‑NUMA VMs outperform dual‑NUMA VMs, so users prefer single‑NUMA packages. However, frequent VM creation and deletion across IDC data centers cause uneven CPU usage, leading to resource fragmentation.

Problem Summary

Problem 1: Repeated VM provisioning and teardown create fragmented resources that cannot be combined into large enough blocks to satisfy high‑CPU or high‑memory VM requests, degrading service capability and overall utilization.

Problem 2: Low overall CPU allocation rates cause VMs to be scattered across multiple compute nodes, leaving individual nodes under‑utilized and increasing operational costs such as power and maintenance.

Dynamic Adjustment Design

The virtualization resource scheduling algorithm allocates CPU, memory, and network resources based on defined policies, aiming to maximize utilization, reduce energy consumption, and maintain service quality. It considers resource availability, demand prediction, load balancing, and priority scheduling.

Two common strategies are Resource Allocation Priority (RAP) and Resource Requirement Priority (RRP), which dynamically monitor VM usage and adjust allocations accordingly.

Module Design

The solution is divided into four functional modules:

Resource Data Collection Module Collects NUMA‑level resource information from compute nodes via Nova API, including NUMA topology, VM placement, VM resource configuration, and bound NUMA node IDs.

Resource Scoring Module Evaluates current resource utilization, fragmentation, and VM distribution to generate a score that determines whether dynamic consolidation is needed.

Consolidation Algorithm Module Executes migration decisions using a resource‑vector minimization algorithm that considers remaining resources, VM demands, and NUMA topology to produce an optimal migration plan.

Live Migration Module Invokes Nova API to perform live VM migrations, records migration details for analysis, and aims for fully automated, non‑disruptive migrations.

Implementation Process

Filter Compute Nodes by Dedicated‑Instance Trait Define a trait for dedicated cloud‑disk instances to ensure only suitable compute nodes are considered, excluding shared‑disk nodes and non‑DPDK nodes.

Identify Dedicated VMs Select VMs that use the dedicated trait for subsequent migration.

Calculate Initial Resource Score (Score1) Score1

Input to Dynamic Consolidation Algorithm Apply the resource‑vector minimization algorithm: sort nodes by remaining resources, migrate VMs from nodes with excess resources to those with deficits, and optionally incorporate load‑balancing or energy‑saving considerations.

Calculate Post‑Consolidation Score (Score2) Score2 is computed after migration; comparing Score2 with Score1 measures optimization effectiveness.

VM Migration Execution Current output is a recommended migration plan requiring manual execution; future work targets fully automated live migration.

Benefits of Dynamic Consolidation

Reduces resource fragmentation, increasing overall utilization and improving the fulfillment rate of large VM packages.

Frees and powers down idle compute nodes, lowering operational costs and energy consumption.

Dynamic consolidation is a key technology for enhancing performance, reducing energy use, and optimizing resource utilization in cloud and data‑center environments, and its continued research will further improve the efficiency and reliability of virtualized systems.