Skip to content

Commit 66f73e5

Browse files
courageJyqlu
andauthored
Add Reinforcement Learning Workload Performance Dashboard (#1226)
Co-authored-by: YQ Lu <yql@google.com>
1 parent f4e0c29 commit 66f73e5

8 files changed

Lines changed: 952 additions & 0 deletions

File tree

Lines changed: 29 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,29 @@
1+
## Reinforcement Learning Workload Performance Dashboard
2+
3+
This dashboard provides comprehensive metrics for monitoring Reinforcement Learning (RL) training workloads. It includes visualizations for loop duration, environment rewards, training loss, model fraction utilization (MFU), and throughput (samples and tokens rates).
4+
5+
### Metrics Collected
6+
7+
The dashboard relies on Prometheus metrics with the prefix `prometheus.googleapis.com/rl.`.
8+
9+
* **Performance**: Loop duration, Training MFU.
10+
* **Training**: Mean Reward, Mean Train Loss.
11+
* **Throughput**: Samples/sec, Tokens/sec (total and per GPU), Total Samples/Episodes/Tokens/Steps rates.
12+
* **Breakdown**: Duration breakdown for Reward, Train, Sample, and Sync phases.
13+
14+
### Prerequisites
15+
16+
* **OTLP Metrics**: Instrument your application or RL framework to record RL metrics using OpenTelemetry (OTLP) following semantic conventions (using the `rl.` prefix). When ingested into Google Cloud, these metrics are made queryable with the `prometheus.googleapis.com/rl.*` prefix.
17+
18+
### Dashboard Deployment
19+
20+
To view this dashboard in your Google Cloud Console:
21+
22+
1. Open the [Google Cloud Monitoring Dashboards](https://console.cloud.google.com/monitoring/dashboards) page.
23+
2. Click **Create Dashboard**.
24+
3. In the top-right corner, click **JSON Editor**.
25+
4. Replace the default contents with the contents of [rl-workload-performance.json](rl-workload-performance.json).
26+
5. Click **Submit** or **Apply**.
27+
28+
---
29+
*Note: This is a sample dashboard and may require adjustments based on your specific metric names or labels if customized.*
Lines changed: 10 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,10 @@
1+
sample_dashboards:
2+
-
3+
category: Reinforcement Learning
4+
id: rl-workload-performance
5+
display_name: "Reinforcement Learning Workload Performance Dashboard"
6+
description: |-
7+
This dashboard provides metrics for generic Reinforcement Learning training workloads, including loop duration, reward, loss, and throughput.
8+
related_integrations:
9+
- id: reinforcement-learning
10+
platform: GKE
339 KB
Loading
407 KB
Loading

0 commit comments

Comments
 (0)