Skip to content

Commit 569a787

Browse files
qiujian16claude
qiujian16
and
claude
committed
docs: add AI workloads section for fleet-level intelligent scheduling
Add a new top-level documentation section for AI workloads that covers OCM's capabilities for managing AI/ML workloads across multiple clusters. The new section includes: - Overview of how OCM helps AI workloads with intelligent placement, resource management, data locality, privacy-preserving training, and lifecycle management - Documentation for MultiKueue integration for AI workload scheduling - Documentation for Fluid integration for fast model loading - ManifestWork workload completion feature for multi-cluster jobs - Dynamic scoring framework for metric-based scheduling - Federated learning controller for privacy-preserving model training - Getting started guide and use cases Fixes #1337 Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com> Signed-off-by: Jian Qiu <jqiu@redhat.com>
1 parent 6b34c0a commit 569a787

3 files changed

Lines changed: 160 additions & 2 deletions

File tree

‎content/en/_index.md‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -35,8 +35,8 @@ The work API that enables resources to be applied to managed clusters from a hub
3535
Dynamic placement of content and behavior across multiple clusters.
3636
{{% /blocks/feature %}}
3737

38-
{{% blocks/feature icon="fa-cloud" title="Vendor Neutral APIs" %}}
39-
Avoid vendor lock-in by using APIs that are not tied to any cloud providers or proprietary platforms.
38+
{{% blocks/feature icon="fa-brain" title="AI Workload Management" url="docs/ai-workloads/" %}}
39+
Run AI workloads across clusters with security and efficiency
4040
{{% /blocks/feature %}}
4141

4242
{{% blocks/feature icon="fa-rocket" title="Launch Apps Everywhere" url="docs/getting-started/integration/app-lifecycle/" %}}

‎content/en/docs/ai-workloads/_index.md‎

Lines changed: 4 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,4 @@
1+
---
2+
title: AI Workloads
3+
weight: 5
4+
---

‎content/en/docs/ai-workloads/fleet-intelligent-scheduling.md‎

Lines changed: 154 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,154 @@
1+
---
2+
title: Fleet-Level Intelligent Scheduling for AI Workloads
3+
weight: 1
4+
---
5+
6+
Open Cluster Management (OCM) provides comprehensive solutions for managing AI workloads across multiple
7+
clusters through intelligent scheduling, efficient model loading, and advanced placement capabilities. This page
8+
lists the key AI integrations and features that help orchestrate AI workloads in multi-cluster environments.
9+
10+
## How OCM Helps AI Workloads on Fleet
11+
12+
AI and machine learning workloads have unique requirements when running at scale across multiple clusters. OCM
13+
addresses these challenges through several key capabilities:
14+
15+
### Intelligent Workload Placement
16+
17+
OCM enables dynamic placement of AI workloads based on cluster capabilities and real-time conditions. You can
18+
schedule training or inference jobs to clusters with available GPU resources, specific hardware accelerators, or
19+
optimal network connectivity. The placement engine considers cluster capacity, resource quotas, and custom scoring
20+
criteria to make intelligent scheduling decisions.
21+
22+
### Resource Management and Queueing
23+
24+
AI workloads often require significant computational resources and benefit from queue-based scheduling. OCM
25+
integrates with workload queueing systems to manage resource allocation across your cluster fleet, ensuring fair
26+
sharing of expensive resources like GPUs while preventing resource contention and enabling priority-based execution.
27+
28+
### Data Locality and Fast Model Loading
29+
30+
Model serving and training workloads require efficient access to large datasets and model files. OCM provides data
31+
caching and acceleration capabilities to reduce model loading time, minimize data transfer costs, and improve
32+
inference latency. This is particularly important in edge computing scenarios where models need to be deployed close
33+
to data sources.
34+
35+
### Privacy-Preserving Distributed Training
36+
37+
For organizations with data distributed across multiple locations or regulatory boundaries, OCM enables federated
38+
learning patterns where training occurs locally on each cluster, and only model updates are shared. This preserves
39+
data privacy and compliance while enabling collaborative model training across the fleet.
40+
41+
### Workload Monitoring and Lifecycle Management
42+
43+
OCM provides unified visibility into AI workload execution across all clusters. You can track job completion status,
44+
monitor resource utilization, and manage the lifecycle of batch training jobs or long-running inference services
45+
from a central control plane.
46+
47+
## AI Integrations and Features
48+
49+
OCM achieves these capabilities through a set of integrations and addons specifically designed for AI workloads.
50+
The following sections detail each integration and how it contributes to the overall AI workload management solution.
51+
52+
## MultiKueue Integration
53+
54+
The MultiKueue integration enables intelligent scheduling of AI workloads across multiple clusters. It leverages
55+
[Kueue](https://kueue.sigs.k8s.io/), a Kubernetes-native job queueing system, to manage and schedule batch
56+
workloads efficiently in a multi-cluster setup.
57+
58+
**Key Features:**
59+
- Queue-based workload management across clusters
60+
- Resource quota enforcement
61+
- Priority-based scheduling
62+
- Fair sharing of cluster resources
63+
64+
**Learn More:**
65+
- [MultiKueue Addon Documentation](
66+
https://github.com/open-cluster-management-io/addon-contrib/blob/main/kueue-addon/README.md)
67+
68+
## Fluid Integration
69+
70+
The Fluid integration provides fast model loading capabilities for AI model serving by leveraging data caching and
71+
acceleration. [Fluid](https://fluid-cloudnative.github.io/) is an open-source Kubernetes-native distributed dataset
72+
orchestrator and accelerator for data-intensive applications.
73+
74+
**Key Features:**
75+
- Accelerated data access for AI models
76+
- Distributed caching of model data
77+
- Support for various storage backends
78+
- Reduced model loading time for inference workloads
79+
80+
**Learn More:**
81+
- [Fluid Addon Documentation](
82+
https://github.com/open-cluster-management-io/addon-contrib/tree/main/fluid-addon)
83+
84+
## ManifestWork for Multi-Cluster Jobs
85+
86+
OCM's ManifestWork API provides a `workload completion` feature that enables easy execution of jobs across multiple
87+
clusters. This feature allows you to track the completion status of workloads deployed to managed clusters.
88+
89+
**Key Features:**
90+
- Deploy and track jobs across multiple clusters
91+
- Monitor workload completion status
92+
- Automatic status aggregation from managed clusters
93+
- Simplified multi-cluster batch job management
94+
95+
**Learn More:**
96+
- [ManifestWork Workload Completion Documentation](
97+
https://open-cluster-management.io/docs/concepts/work-distribution/manifestwork/#workload-completion)
98+
99+
## Dynamic Scoring Framework
100+
101+
The dynamic scoring addon enables intelligent workload scheduling based on real-time metrics and custom scoring
102+
algorithms. This framework allows you to implement sophisticated placement decisions based on cluster conditions,
103+
resource availability, and custom business logic.
104+
105+
**Key Features:**
106+
- Metric-based cluster scoring
107+
- Custom scoring algorithms
108+
- Dynamic placement decisions
109+
- Real-time resource awareness
110+
111+
**Learn More:**
112+
- [Dynamic Scoring Framework Documentation](
113+
https://github.com/open-cluster-management-io/addon-contrib/blob/main/dynamic-scoring-framework/README.md)
114+
115+
## Federated Learning Controller
116+
117+
The federated learning controller enables secure model training across multiple clusters while preserving data
118+
privacy. This addon implements federated learning patterns where data remains in its original location, and only
119+
model updates are shared across clusters.
120+
121+
**Key Features:**
122+
- Privacy-preserving model training
123+
- Distributed learning across clusters
124+
- Data locality and security
125+
- Aggregation of model updates
126+
127+
**Learn More:**
128+
- [Federated Learning Controller Documentation](
129+
https://github.com/open-cluster-management-io/addon-contrib/blob/main/federated-learning-controller/README.md)
130+
131+
## Getting Started
132+
133+
To begin using these AI workload capabilities:
134+
135+
1. Ensure you have a working OCM environment with at least one managed cluster. See the
136+
[Getting Started Guide](/docs/getting-started/) for initial setup.
137+
138+
2. Install the specific addons you need from the
139+
[addon-contrib repository](https://github.com/open-cluster-management-io/addon-contrib).
140+
141+
3. Configure placement policies to target the appropriate clusters for your AI workloads.
142+
143+
4. Deploy your AI workloads using the appropriate integration method.
144+
145+
## Use Cases
146+
147+
These AI integrations enable various multi-cluster AI scenarios:
148+
149+
- **Distributed Training**: Train large models across multiple clusters with data locality
150+
- **Batch Inference**: Run inference jobs at scale across your fleet
151+
- **Model Serving**: Deploy and serve models with optimized data loading
152+
- **Resource Optimization**: Dynamically place workloads based on GPU availability and cluster metrics
153+
- **Privacy-Preserving ML**: Train models on distributed data without centralizing sensitive information
154+
- **Edge AI**: Deploy and manage AI workloads at the edge with efficient model distribution and local inference

0 commit comments

Comments
 (0)