Operating System Console

更新时间:
复制 MD 格式

The Operating System Console is Alibaba Cloud's official intelligent platform for operating system management and diagnostics. It allows O&M and development teams to visually troubleshoot and manage operating systems, eliminating the need to log in to servers and use multiple command-line tools. You can manage the console through APIs, SDKs, CLIs, and MCP.

Use cases

  • Daily O&M:

    • Use System Health to quickly identify server performance bottlenecks.

    • Use System Diagnosis to automatically analyze complex issues, such as out-of-memory (OOM) process kills, I/O spikes, network jitter, and insufficient memory.

    • Use Crash Diagnosis to automatically interpret panic logs and receive repair suggestions.

  • AI/GPU scenarios:

    • AI Profiling non-intrusively collects full-stack call traces to determine if communication or operators are causing training bottlenecks.

    • If distributed training slows down due to specific nodes, use GPU Diagnosis to quickly identify the slow nodes.

    • If training hangs, GPU Diagnosis automatically determines whether the issue is a hardware failure or a communication problem.

  • Simplified troubleshooting: The OS Copilot allows you to use natural language to troubleshoot kernel-level issues and receive expert-level diagnostic advice.

Region restrictions

This feature is currently available only in the Chinese mainland and China (Hong Kong).

Prerequisites

  • If you are a RAM user, you must have the AliyunECSReadOnlyAccess, AliyunSubManageFullAccess, and AliyunSysomFullAccess policies.

  • When you log on to the Operating System Console for the first time, click Activate Service to enable console permissions.

Quick start

  1. Log on to the Operating System Console.

  2. In Resource Management > Component Management, install the SysOM component on the target ECS instance.

  3. In Resource Management, confirm that the instance status is Managed.

  4. Go to Overview or other feature modules to use the corresponding features.

Note: You can use some diagnostic features on Unmanaged instances without managing them first.

Features

The navigation pane on the left of the console contains the following feature modules:

Top-level menu

Second-level menu

Description

Overview

Provides a global view of resources and quick-access entry points.

System Health & Diagnosis

System Health

Monitors the health of clusters and nodes.

System Diagnosis

Node Diagnosis, Pod Diagnosis, and Historical Diagnosis.

Process Hotspot Analysis

Tracks process-level performance hotspots and generates flame graphs.

Hotspot Comparison Analysis

Compares hotspots between different points in time or different instances.

GPU Performance & Diagnosis (Invitation-only Beta)

GPU Diagnosis

Diagnoses GPU anomalies in AI workloads.

AI Profiling

Provides full-lifecycle performance observability for AI applications.

Continuous GPU Profiling

Continuously analyzes CPU and GPU performance hotspots.

Resource Management

Resource Management

Manages instances and clusters.

Component Management

Installs, upgrades, and uninstalls extension components such as SysOM.

Abnormal Event Alarms

Abnormal Event List

Queries and filters abnormal events.

Policy Management

Creates and manages alarm policies.

Subscription Management

Provides subscription services for operating system security updates.

In the upper-right corner of the console, you can also access the OS Copilot intelligent assistant dialog box.

Overview

The Overview page provides a high-level view of your resources and overall system health.

Section

Description

Resource Management Status

Displays the number of managed clusters and nodes.

Abnormal Event Dashboard

Displays the current number of abnormal events across six dimensions: cluster, node, disk traffic, memory, scheduling, and network.

Event Alarms

Summarizes recent abnormal event alarms.

Subscription Services

Displays your current subscriptions, such as the Anolis 7 security update subscription.

System health and diagnosis

System health

System health reflects the overall health of a cluster, node, or container, based on key monitoring metrics.

Key features:

Feature

Description

Cluster health score

Allows you to select a managed cluster and view its health score and health metrics to assess its overall condition.

Node count and resource monitoring

Displays node distribution and key resource metrics, such as CPU (utilization rate, allocation rate), memory (utilization rate, allocation rate), and disk (I/O utilization rate, allocation rate).

Historical health trend

View health trends over the last 24 hours, week, or month.

Abnormal event analysis

Displays the distribution and details of abnormal events, with support for multi-dimensional filtering. It integrates a large language model to provide context-aware root cause analysis and repair suggestions for alarms.

System diagnosis

System Diagnosis provides multi-dimensional operating system diagnostic capabilities to help you quickly identify and analyze system issues. Use the drop-down list at the top to switch between three diagnosis modes:

Diagnosis modes:

Mode

Description

Node Diagnosis

Performs system diagnosis on a single node. Multiple diagnosis items are supported.

Pod Diagnosis

Diagnoses Pods managed by ACK and ACS clusters. This mode is suitable for containerized scenarios.

Historical Diagnosis

Performs retrospective analysis of historical issues, such as OOM events, on ACS and ECI instances.

Diagnosis items (Node Diagnosis mode):

Diagnosis type

Diagnosis item

Description

Architectures

Operating systems

Memory Diagnosis

Panoramic memory analysis

Provides a panoramic view of system memory. It supports Glibc local memory allocation output, Java Native Memory analysis, and detection of tcp_mem and socket leaks. Java memory diagnosis supports large file uploads and generates on-heap and off-heap flame graphs.

x86, ARM64

  • Alibaba Cloud Linux 2/3/3 Pro

  • CentOS 7/8

  • Rocky Linux 8.8 and later

  • Ubuntu 22.04/24.04

  • Anolis OS 7/8

OOM diagnosis

Diagnoses and analyzes OOM events with enhanced data collection.

Storage Diagnosis

I/O traffic analysis

Analyzes the distribution and usage of system I/O traffic.

One-click I/O diagnosis

Performs one-click rapid diagnosis for frequently occurring issues such as high I/O latency, I/O bursts, and high I/O wait times.

Network Diagnosis

Packet loss diagnosis

Diagnoses the causes of network packet loss.

Network jitter

Analyzes network jitter issues caused by slow packet reception, softirqs, and scheduler problems.

Scheduling Diagnosis

Scheduling jitter diagnosis

Diagnoses system scheduling jitter issues.

System load diagnosis

Analyzes the causes of an abnormal 1-minute average system load (load1 metric).

Scenario Diagnosis

Crash diagnosis

Analyzes crash causes and provides repair suggestions. Supports hardlock and hung task analysis.

One-Click Diagnosis

One-click diagnosis

Performs a one-click scan for common OS issues and generates a diagnostic report.

Process hotspot analysis

Analyzes hotspots on a single node over a specific period, generating hotspot time-series charts, flame graphs, and call graphs. It supports hotspot tracking for Java processes (by collecting Java stack information) and querying historical hotspot records.

Item

Description

Supported architectures

x86

Supported operating systems

Alibaba Cloud Linux 2/3/3 Pro, CentOS 7/8, Rocky Linux 8.8 and later, Ubuntu 22.04/24.04, and Anolis OS 7/8

Hotspot comparison analysis

Compares hotspots between different points in time or different instances to help identify the causes of performance changes.

Item

Description

Supported architectures

x86

Supported operating systems

Alibaba Cloud Linux 2/3/3 Pro, CentOS 7/8, Rocky Linux 8.8 and later, Ubuntu 22.04/24.04, and Anolis OS 7/8

GPU performance and diagnosis (invitation-only beta)

The GPU Performance & Diagnosis module is designed for AI workloads and provides capabilities for GPU-level performance observation, diagnosis, and optimization. This is an advanced feature.

GPU diagnosis

Diagnoses system anomalies in AI workloads and generates diagnostic conclusions, GPU operational status, and AI job status.

Item

Description

Fault diagnosis

Diagnoses inference service or training anomalies (task hangs, inference latency), GPU hardware anomalies (card drops, XID errors), and NCCL anomalies (network hangs, operator hangs), and locates slow nodes in large language model training.

Slow node diagnosis

Provides slow node diagnosis for large language model training scenarios. The diagnostic report includes basic information, analysis conclusions, and raw data visualized with Perfetto.

  • Supported architecture: x86

  • Supported operating systems: Alibaba Cloud Linux 2/3, Ubuntu 22.04/24.04

  • Other requirements: Available only for GPU instances. Requires specific NCCL versions (v2.21.5.1 to v2.28.9.1).

AI profiling

An advanced analysis tool for observing, diagnosing, and optimizing the performance of AI applications throughout their lifecycle. It provides end-to-end performance analysis by tracing cross-layer software stack calls during AI model training and inference. It is non-intrusive and requires no container modifications.

Item

Description

Tracing scope

Python stack, Torch layer, GPU memory, CudaRuntime, and GPU kernel functions.

Collection modes

Duration mode (collects data for a specified time, from 1,000 ms to 5,000 ms) and Iteration mode (collects data by iteration, with an option to skip the first n iterations).

Analysis results

Provides analysis suggestions, CPU/GPU summaries, GPU kernel analysis (including Tensor Cores usage), iteration statistics with differential analysis, and CPU/GPU tracing visualization, which includes a built-in TimeLine view. Diagnostic reports can be exported for offline viewing.

Supported architecture

x86

Supported operating systems

Alibaba Cloud Linux 2/3, Ubuntu 22.04/24.04

Other requirements

Python 3.9 to 3.12, torch 2.4 to 2.7, and CUDA 12.0 to 12.8 (excluding 12.7). Available only for GPU instances (A-series, L-series, and T-series cards). The target process must use the GPU and consume at least 0.5 GB of memory per second or per iteration.

Continuous GPU profiling

Helps analyze performance hotspots of AI applications on both CPUs and GPUs. It visualizes function call stacks and time consumption distribution to locate performance bottlenecks and optimize AI task execution efficiency.

Item

Description

Analysis views

CPU/GPU heatmap (each column represents 1 second and contains 50 cells, with each cell representing 20 ms), CPU flame graph (hotspot graph of process function call relationships), and GPU flame graph (displays GPU call stack information related to the Python process).

Prerequisites

Install or upgrade the SysOM component to version 3.9.0 or later. Enabling this feature increases the memory limit of the SysOM Agent from the default of 300 MB to 2 GB.

Supported architecture

x86

Supported operating systems

Alibaba Cloud Linux 2/3, Ubuntu 22.04/24.04


Resource management

Resource management

The Resource Management page lets you manage your instances and clusters. The page has three tabs:

Tab

Description

Managed

Displays a list of managed instances, including resource management expiration time, instance ID/name, health score, image type, SysOM component version, and SysOM component configuration, and allows you to perform canary deployments.

Unmanaged

Lists unmanaged instances that you can choose to manage.

Clusters

Displays information about managed clusters, allows you to manage ACK clusters and configure auto scaling for their instances.

Component management

Manages the entire lifecycle of OS extension components such as SysOM, including installation, upgrades, and uninstallation.

Feature

Description

Install/upgrade/uninstall components

Manages extension components such as SysOM. Supports Rocky Linux 8.8 and later.

Canary deployment

Supports canary deployment of components to ACK clusters based on the number of nodes, percentage of nodes, or ACK node labels.

Configuration management

Supports enabling and configuring features, including the FastOOM feature (monitors node-level memory pressure and triggers a node-level FastOOM process-killing function with regex match details) and the zombie memcg reclamation feature.

Abnormal event alarms

The Abnormal Event Alarms module provides capabilities for viewing and filtering abnormal events and managing alarm policies.

Abnormal event list

Displays abnormal events detected in the system. You can filter events by time range (last 24 hours, last 3 days, last week, last month, or a custom range) and various other conditions.

The list displays the following information: abnormal event type, description, severity level (Critical/Warning/Info), detection time, and associated cluster name or instance ID.

Supported abnormal event types include abnormal CPU utilization, tcp_mem anomaly, socket leak, abnormal I/O read/write latency, GPU anomaly, and crash event. Alarm information supports push notifications for GPU anomalies and can display Kubernetes label information.

Policy management

Creates and manages alarm policies for abnormal events.

Feature

Description

Create policy

Creates a new alarm policy by configuring alarm conditions and notification methods. Notifications can be sent through multiple channels, such as email, SMS, and instant messaging tools.

Policy list

Manages created policies. Displays information such as policy name, number of events, cluster name, and whether the policy is active.

Subscription management

Subscription Management provides security update subscription services for operating systems that have reached their end-of-life (EOL).

Subscription service

Description

CentOS 7 security update subscription

CentOS 7 reached its EOL on June 30, 2024. You can obtain security updates through this subscription service.

Alibaba Cloud Linux 2 Extended Lifecycle Support (ELS)

Alibaba Cloud Linux 2 reached its EOL on March 31, 2024. You can continue to receive support through the ELS service.

Anolis 7 security update subscription

Anolis OS 7 reached its EOL on June 30, 2024. You can obtain security updates through this subscription service.

Note: You can unsubscribe from CentOS security update and Alibaba Cloud Linux 2 ELS subscriptions in the console.

OS Copilot

OS Copilot is Alibaba Cloud’s self-developed Linux operating system intelligent assistant. It is available in the dialog box in the upper-right corner of the console and supports natural language Q&A, assisted command execution, script/code generation, and system operations tuning.

Core capabilities:

Capability

Description

Professional OS Q&A

Provides expert answers to OS-related questions and can include links to documentation in its responses.

Assisted command execution

Helps generate and execute Linux commands based on natural language descriptions.

Script/code generation

Automatically generates O&M scripts or code based on your requirements.

Scenario-based tool integration

Integrates with scenario-based tools such as System Diagnosis to provide a one-stop O&M experience.

Intelligent diagnosis agent

Includes a new high CPU load diagnosis feature that automatically identifies and analyzes sudden increases in CPU usage to quickly locate hotspot threads or methods. It also guides users through the diagnosis of anomalies such as OOM events by assisting with information collection and automatically running the panoramic memory analysis tool.

Diagnosis MCP service

Supports the Diagnosis MCP service, allowing third-party intelligent agents to access the console's diagnostic capabilities through MCP.

How to use:

  • Linux command line: Use it directly on a server from the command line.

  • Console dialog box: Start a conversation from the dialog box in the upper-right corner of the console.