The EAS monitoring dashboard is a unified panel for tracking real-time performance across all deployed EAS services in a region — without checking each service individually.
Dashboard data is region-scoped and aggregates metrics from all services across all workspaces in the current region under your Alibaba Cloud account. Switching the workspace only changes the console context and does not affect the dashboard's data scope.
Use cases
-
Overall service overview: Assess the health of all EAS services in the current region at a glance.
-
Performance monitoring: Track key aggregated metrics in real time — QPS, response time, and replica count.
-
Resource management: Monitor CPU, memory, and GPU usage to inform scale-out and scale-in decisions.
-
Troubleshooting: Identify performance anomalies and service issues from aggregated monitoring data.
-
Capacity planning: Use historical monitoring data to plan future capacity.
View the dashboard
-
Log on to the PAI console. Select a region at the top of the page, then select the desired workspace and click Elastic Algorithm Service (EAS).
-
Switch to the Monitoring tab. The dashboard supports the following interactions:
-
Time range selection: Select a preset range such as the last 5 minutes or 15 minutes. The default is the last 6 hours.
-
Manual refresh: Click the refresh button to get the latest data, or set an auto-refresh frequency.
-
Dimension filtering: Filter data by dimension — for example, filter by the User dimension to view data for all users or a specific user.
-
Chart interaction: Hover over a chart to inspect data points, or click a legend entry to toggle its display.
-
Monitoring metrics
Service dashboard
The following table describes all metrics on the Service dashboard. Calculations apply to all EAS service instances in the current region under your Alibaba Cloud account. Sum is the total value across all instances; Average is the mean.
QPS and Avg RT are the most important performance indicators. Replicas reflects cluster scaling activity. The CPU, memory, and GPU metrics reflect resource consumption. Traffic and Daily Invoke reflect overall traffic and call volume.
| Category | Metric | Description | Calculation | Purpose |
|---|---|---|---|---|
| Request performance | QPS | Requests per second. | Sum | Evaluate overall cluster throughput; identify traffic peaks and fault storms when combined with error rate data. |
| Avg RT | Average response time — the average time from sending a request to receiving a response. | Average | Evaluate overall cluster response speed; identify global performance bottlenecks. | |
| Instance scale | Replicas | Instance count. Includes three sub-metrics: TotalReplicas (total), PendingReplicas (standby), and Available Replicas (active). | Sum | Understand real-time cluster scale; monitor whether auto scaling behavior meets expectations. |
| Replicas By Resource | Instance count broken down by resource type. | Sum | Understand resource distribution across instance types. | |
| CPU | CPU Total | Total CPUs available to the service. | Sum | Evaluate the total reserved CPU capacity. |
| CPU Utilization | CPU utilization percentage. | Average | Evaluate overall cluster CPU efficiency for capacity planning and cost optimization. | |
| CPU Usage | CPU usage. | Average | Estimate the cost of consumed CPU resources. | |
| Memory | Memory | Memory usage. Includes RSS (sum of resident memory) and Cache (sum of page cache). | Average | Monitor system memory usage to determine resource pressure and locate performance bottlenecks. |
| Memory Utilization | Memory utilization percentage. | Average | Evaluate overall memory resource efficiency. | |
| GPU | GPU Total | Total GPUs in use. | Sum | Evaluate GPU resource scale. |
| GPU Utilization | GPU utilization percentage. | Average | Evaluate GPU resource efficiency. | |
| GPU Memory | Actual GPU memory usage. | Average | Evaluate GPU memory consumption. | |
| Traffic and calls | Traffic | Sum of inbound (In) and outbound (Out) network traffic for all services. | Sum | Reflects network communication status across services. |
| Daily Invoke | Daily call count, categorized by HTTP status code. | Sum | Observe long-term business health and error rate trends. |
GPU dashboard
The GPU dashboard provides detailed monitoring for GPU-based services to help optimize GPU resource utilization.
The GPU dashboard is especially useful for identifying cost inefficiencies. Focus on the low-utilization metrics to find services with wasteful resource configurations.
| Category | Metric | Description | Purpose |
|---|---|---|---|
| Overview | Total GPU usage | Total GPUs used by all services in the current region. | Indicates overall GPU scale for capacity planning. |
| Average GPU utilization | Average GPU utilization percentage across all services that use GPUs. | Measures overall GPU efficiency; identifies risk of resource waste. | |
| Number of services using GPUs | Total number of services currently using GPU resources. | Understand GPU service distribution and breadth of resource usage. | |
| Utilization by service | Number of services with average GPU utilization below 10% | Services whose average GPU utilization is below 10%. | Identifies severe resource waste; prioritize these services for optimization or release. |
| Number of services with average GPU utilization below 30% | Services whose average GPU utilization is below 30%. | Identifies low-utilization services; consider adjusting their resource configurations. | |
| Number of services with average GPU utilization above 50% | Services whose average GPU utilization is above 50%. | Confirms effective GPU utilization; use to verify optimization results. | |
| GPU utilization distribution | Number of GPUs with average utilization below 10% | GPU cards whose average utilization is below 10%. | Identifies idle GPUs that can be released to reduce costs. |
| Number of GPUs with average utilization below 30% | GPU cards whose average utilization is below 30%. | Identifies low-utilization GPUs; adjust resource configurations. | |
| Number of GPUs with average utilization above 50% | GPU cards whose average utilization is above 50%. | Confirms fully utilized GPUs; evaluate whether to scale out. | |
| Resource distribution | Number of GPUs in dedicated resource groups (including Lingjun) | GPUs in dedicated resource groups, including Lingjun. | Monitors dedicated resource changes for capacity planning. |
| Number of GPUs in public resource groups | GPUs in public resource groups. | Monitors public resource usage to optimize allocation policies. | |
| Number of GPUs in spot instances (including Lingjun) | GPUs in spot instances, including Lingjun. | Monitors low-cost GPU resources to help balance cost and stability. | |
| Detailed utilization | Detailed GPU utilization per service | Per-service GPU utilization. Displays "No data" when no data is available. | Enables in-depth analysis of individual service GPU usage to locate performance issues. |
Usage notes
-
Dashboard data is aggregated from all services in the region and may be delayed.
-
For detailed monitoring on a specific service, open that service's monitoring page.
FAQ
Why is the CPU utilization on the dashboard low, but my service fails to scale out or reports insufficient resources?
The dashboard shows the average CPU utilization across all services. Idle services pull this average down, even when the cluster's physical resources are fully exhausted.
Check the Pending_Replicas metric. If it stays consistently above 0, the cluster resource pool is full and cannot schedule new instances. Scale out the cluster or optimize the resource configurations of existing services.
How do I view detailed monitoring for a specific service?
Go to the Inference Service tab, select the target service, and open its details page. Switch to the Monitoring tab to view detailed charts. For metric descriptions, see Service monitoring overview.