CNCF Project
KubeEdge
Term
2026 Term 3 (Sep-Nov)
Program Name
KubeEdge-Ianvs Simulation Sandbox: Environment-Isolated Execution
Program Description
Description
For most distributed AI scheme developers, establishing and deploying a large-scale cloud-edge collaborative system is often complicated and burdensome. While KubeEdge-Ianvs currently provides a single-node algorithm tester to evaluate accuracy-wise metrics using test datasets, measuring system-wise metrics, such as bandwidth, computing power, and peak memory, in a real-world manner for large-scale nodes is extremely difficult and costly. Furthermore, executing all test cases within a single shared Python process easily triggers dependency conflicts, path contamination, and fatal OOM crashes when heavy-load models like LLMs, VLAs, and foundation models run alongside lightweight examples. To address these challenges, this project would like to introduce an industrial-grade distributed collaborative system simulation using a worker-in-worker approach on a single machine, providing low-cost, scalable testing capabilities, robust environment isolation, and precise system-level metric profiling.
Expected Outcome
Restoration of Simulation Core Functions: Considering previous proposal Restore and extend the Ianvs 2022 simulation proposal (Ianvs PR #35) and implementation (Ianvs PR #39, with 5+ identified breakages), to build the Simulation Controller to isolate each test case within an independent transient runtime environment, and enforce system-level resource quotas and boundary control mechanisms to restrict edge node resources (CPU and memory), avoiding dependency conflicts and OOM risks. Key components include:
Environment Administrator of Simulation Controller: Introduce a simulation controller within the test case controller to provide a worker-in-worker system on a single machine, simulating multi-node systems. Implement the Simulation Environment Administrator to parse system configurations, check host environment requirements (e.g., memory > 4GB), and automatically build, deploy, close, and delete the simulation environment.
Simulation Job Administrator of Simulation Controller: Develop the critical Simulation Job Administrator to handle algorithm image building (e.g., Docker), YAML generation, job deployment/deletion, and list-watching of simulation results with workers. Simultaneously, deploy an isolated execution layer using transient runtime environments and system resource quotas (CPU and memory) to completely prevent dependency conflicts and OOM crashes.
Verify Cluster with Multi-Dimensional Metrics: Complete KubeEdge-native cluster simulation validation using kind + edgecore + Sedna all-in-one scripts.
Multi-Dimensional Metrics Integration: Align underlying system metrics (CPU utilization, peak memory, wall-clock time) with upper-layer algorithm metrics, presenting them uniformly in the existing StoryManager leaderboard to achieve end-to-end comprehensive performance evaluation for distributed AI.
Technologies
KubeEdge-Ianvs, KubeEdge-Sedna, KubeEdge, Kubernetes, Docker, Linux Kernel mechanisms, Go, Python, Benchmark, AI/ML
Skills same as Technologies?
Required/Desirable Skills
No response
Mentors
Zimu Zheng | @MooreZheng | zimu.zheng@huawei.com | moore.zheng
Shijing Hu | @hsj576 | sjhu21@m.fudan.edu.cn | hsj576
Upstream Issue URL
kubeedge/ianvs#348
Application Prerequisites
Coding Challenge URL
No response
Custom Prerequisite Name
No response
Custom Prerequisite Description
No response
Custom Prerequisite — File Upload
CNCF Project
KubeEdge
Term
2026 Term 3 (Sep-Nov)
Program Name
KubeEdge-Ianvs Simulation Sandbox: Environment-Isolated Execution
Program Description
Description
For most distributed AI scheme developers, establishing and deploying a large-scale cloud-edge collaborative system is often complicated and burdensome. While KubeEdge-Ianvs currently provides a single-node algorithm tester to evaluate accuracy-wise metrics using test datasets, measuring system-wise metrics, such as bandwidth, computing power, and peak memory, in a real-world manner for large-scale nodes is extremely difficult and costly. Furthermore, executing all test cases within a single shared Python process easily triggers dependency conflicts, path contamination, and fatal OOM crashes when heavy-load models like LLMs, VLAs, and foundation models run alongside lightweight examples. To address these challenges, this project would like to introduce an industrial-grade distributed collaborative system simulation using a worker-in-worker approach on a single machine, providing low-cost, scalable testing capabilities, robust environment isolation, and precise system-level metric profiling.
Expected Outcome
Restoration of Simulation Core Functions: Considering previous proposal Restore and extend the Ianvs 2022 simulation proposal (Ianvs PR #35) and implementation (Ianvs PR #39, with 5+ identified breakages), to build the Simulation Controller to isolate each test case within an independent transient runtime environment, and enforce system-level resource quotas and boundary control mechanisms to restrict edge node resources (CPU and memory), avoiding dependency conflicts and OOM risks. Key components include:
Environment Administrator of Simulation Controller: Introduce a simulation controller within the test case controller to provide a worker-in-worker system on a single machine, simulating multi-node systems. Implement the Simulation Environment Administrator to parse system configurations, check host environment requirements (e.g., memory > 4GB), and automatically build, deploy, close, and delete the simulation environment.
Simulation Job Administrator of Simulation Controller: Develop the critical Simulation Job Administrator to handle algorithm image building (e.g., Docker), YAML generation, job deployment/deletion, and list-watching of simulation results with workers. Simultaneously, deploy an isolated execution layer using transient runtime environments and system resource quotas (CPU and memory) to completely prevent dependency conflicts and OOM crashes.
Verify Cluster with Multi-Dimensional Metrics: Complete KubeEdge-native cluster simulation validation using kind + edgecore + Sedna all-in-one scripts.
Multi-Dimensional Metrics Integration: Align underlying system metrics (CPU utilization, peak memory, wall-clock time) with upper-layer algorithm metrics, presenting them uniformly in the existing StoryManager leaderboard to achieve end-to-end comprehensive performance evaluation for distributed AI.
Technologies
KubeEdge-Ianvs, KubeEdge-Sedna, KubeEdge, Kubernetes, Docker, Linux Kernel mechanisms, Go, Python, Benchmark, AI/ML
Skills same as Technologies?
Required/Desirable Skills
No response
Mentors
Zimu Zheng | @MooreZheng | zimu.zheng@huawei.com | moore.zheng
Shijing Hu | @hsj576 | sjhu21@m.fudan.edu.cn | hsj576
Upstream Issue URL
kubeedge/ianvs#348
Application Prerequisites
Coding Challenge URL
No response
Custom Prerequisite Name
No response
Custom Prerequisite Description
No response
Custom Prerequisite — File Upload