gpu-recommendation

command
v0.6.4 Latest Latest
Warning

This package is not in the latest version of its module.

Go to latest
Published: Nov 6, 2026 License: Apache-2.0 Imports: 18 Imported by: 0

README

CM-Beetle GPU Recommendation Test CLI

Automated testing tool for evaluating CM-Beetle's GPU-accelerated infrastructure recommendation and 6-dimension match rate evaluation without provisioning actual cloud VMs.


1. Sample Source Workload Specification (Input)

The default test scenario uses an on-premise AI inference / lightweight training server equipped with an NVIDIA Tesla T4 GPU:

Category Specification Details
Server ID nvidiat4-1 Standalone AI workload node
CPU 4 vCPUs 1 Socket, 2 Cores, 4 Threads (Architecture: x86_64)
Memory 16 GiB DDR4 RAM
Root Disk 100 GB SSD
Operating System Ubuntu 22.04.4 LTS Codename: jammy, Kernel Linux 5.15+
GPU Accelerator 1x NVIDIA Tesla T4 16 GB VRAM, Driver: 570.195.03, CUDA: 12.8

2. Sample Recommendation Results by Target CSP (Output)

The following table summarizes the actual recommended VM specs, GPU configurations, OS images, and multi-dimensional match rates evaluated by CM-Beetle across the supported CSPs for the input source workload (nvidiat4-1: 4 vCPU, 16 GiB RAM, 1x Tesla T4 16GB, Ubuntu 22.04):

Target CSP / Region Recommended Node Spec Recommended OS Image Overall Match Rate Status Notes
AWS
ap-northeast-2
g4dn.xlarge
4 vCPU, 16 GiB RAM
GPU: 1x NVIDIA Tesla T4 (16 GB)
(aws+ap-northeast-2+g4dn.xlarge)
ami-012a353bb3afb92ee
(Canonical Ubuntu Server 22.04 LTS Base)
Min: 100.0%
Max: 100.0%
Avg: 100.0%
highly-matched Exact 100% Match across all 6D
CPU, Mem, Image (Clean Base), GPUCount, GPUVRAM, GPUModel all 100% exact match.
GCP
asia-northeast3
g2-standard-4
4 vCPU, 16 GiB RAM
GPU: 1x NVIDIA L4 (24 GB)
(gcp+asia-northeast3+g2-standard-4)
ubuntu-2204-jammy-v20260820
(Google Cloud Ubuntu 22.04 LTS Base)
Min: 75.0%
Max: 100.0%
Avg: 92.7%
partially-matched L4 Match with Clean Base OS
CPU 100%, Mem 97.7%, Image 83.3%, GPUCount 100%, VRAM 100%, GPUModel 75% (L4 vs T4).
NCP
kr
gp1t4-g1-h50
4 vCPU, 16 GiB RAM
GPU: 1x NVIDIA Tesla T4 (16 GB)
(ncp+kr+gp1t4-g1-h50)
23214590
(Ubuntu 22.04 Standard Base)
Min: 66.7%
Max: 100.0%
Avg: 91.1%
partially-matched Native T4 Match with Clean Base OS
CPU 100%, Mem 80.0%, Image 66.7%, GPUCount 100%, GPUVRAM 100%, GPUModel 100%.
NHN Cloud
kr1
g2.t4.c4m32
4 vCPU, 32 GiB RAM
GPU: 1x NVIDIA Tesla T4 (16 GB)
(nhn+kr1+g2.t4.c4m32)
0f07c795-2a46-44fc-a61b-fa0d96763ce2
(Ubuntu Server 22.04 LTS Base)
Min: 50.0%
Max: 100.0%
Avg: 86.1%
partially-matched Native T4 Match with Clean Base OS
CPU 100%, Mem 50% (32GB bundle), Image 66.7%, GPUCount 100%, GPUVRAM 100%, GPUModel 100%.
Alibaba Cloud
ap-northeast-2
ecs.gn7i-c8g1.2xlarge
8 vCPU, 32 GiB RAM
GPU: 1x NVIDIA A10 (24 GB)
(alibaba+ap-northeast-2+ecs.gn7i-c8g1.2xlarge)
ubuntu_22_04_x64_20G_alibase_20260810.vhd
(Alibaba Ubuntu 22.04 Base)
Min: 50.0%
Max: 100.0%
Avg: 74.2%
partially-matched A10 Match with Clean Base OS
8 vCPU, 32GB RAM, 1x A10 24GB, Image 66.7%, GPUCount 100%, GPUVRAM 100%, GPUModel 75%.
Azure
koreasouth
standard_nd96isr_h200_v5
96 vCPU, 1900 GiB RAM
GPU: 8x NVIDIA H200 (141 GB)
(azure+koreasouth+standard_nd96isr_h200_v5)
Canonical:0001-com-ubuntu-server-jammy:22_04-lts-gen2:22.04.202608060
(Canonical Ubuntu Server 22.04)
Min: 0.9%
Max: 100.0%
Avg: 46.0%
partially-matched Clean Base OS with Available GPU
Canonical official base image selected for target GPU spec in region.
Tencent Cloud
ap-seoul
N/A N/A 0.0% nothing-to-recommend No compatible GPU spec and OS image pair registered in the local catalog.
IBM Cloud
au-syd
N/A N/A 0.0% nothing-to-recommend No compatible GPU spec and OS image pair registered in the local catalog.

[!NOTE]

  • Evaluation Dimensions (6D): CPU, Memory, Image, GPUCount, GPUVRAM, GPUModel.
  • Status Classification:
    • highly-matched: All active dimensions meet or exceed minMatchRate (default: 85.0%).
    • partially-matched: Valid candidate where some dimensions deviate (e.g., higher memory bundle or different GPU model tier).
    • nothing-to-recommend: No compatible spec-image pairs found in target region.

3. Quick Start

Run recommendation tests across all configured CSPs:

# Execute from the cm-beetle root directory
make test-gpu-recommendation

Or run directly from this directory:

cd cmd/test-cli/gpu-recommendation
go run main.go -config testconf/test-config.yaml

4. Configuration

testconf/test-config.yaml:

test:
  set:
    mode: sequential # sequential or parallel
  cases:
    - csp: aws
      region: ap-northeast-2
      name: AWS-Seoul
      execute: true
    - csp: azure
      region: koreasouth
      name: Azure-Busan
      execute: true
    - csp: gcp
      region: asia-northeast3
      name: GCP-Seoul
      execute: true
    - csp: alibaba
      region: ap-northeast-2
      name: Alibaba-Seoul
      execute: true
    - csp: tencent
      region: ap-seoul
      name: Tencent-Seoul
      execute: true
    - csp: ibm
      region: au-syd
      name: IBMCloud-Sydney
      execute: true
    - csp: ncp
      region: kr
      name: NCP-Seoul
      execute: true
    - csp: nhn
      region: kr1
      name: NHNCloud-Pangyo
      execute: true

beetle:
  endpoint: http://localhost:8056
  requestBodyFile: testconf/source-infra-gpu-t4.json
  authConfigFile: testconf/auth-config.json

5. Output Reports

  • Console: Displays real-time progress, candidate details, and 6-dimension scores.
  • Markdown Report: Written to testresult/gpu-recommendation-test-report.md.

6. Architecture Strategy: OS Image Matching & GPU Stack Setup

  1. Standard Base OS Image Priority:
    • Standard clean base OS images (e.g. Canonical Ubuntu 22.04 LTS) are prioritized as the primary recommendation target across all CSPs.
    • Eliminates artificial non-GPU image penalties, preserving genuine 100% OS similarity for matching base distributions.
    • Prevents catalog search failures (nothing-to-recommend) in CSPs or regions where pre-baked GPU images are not registered.
  2. GPU Image as Graceful Fallback:
    • Pre-baked GPU images (IsGPUImage=true) are relegated to a secondary fallback when no basic OS image is registered in the target catalog.
  3. Roadmap: Idempotent GPU Node Bootstrap Script & API:
    • Provide an idempotent initialization script and dedicated API (POST /beetle/.../gpu-bootstrap) that detects, configures, and verifies the full accelerator stack:
      • NVIDIA Kernel Driver
      • CUDA Toolkit and Runtime
      • NVIDIA Container Toolkit (nvidia-docker2 runtime)
      • Fabric Manager (for multi-GPU NVLink nodes)
    • Guarantees a predictable, reproducible execution environment regardless of whether the node was booted from a base OS or a pre-baked image.

Documentation

Overview

Package main provides the CM-Beetle GPU Recommendation Test CLI.

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL