This document explores how VMware vSphere 8 Update 2 optimizes VM placement on hosts with GPUs to improve hardware utilization. The focus is on vGPU consolidation, ensuring that VMs with similar GPU profiles are placed efficiently to maximize hardware use and reduce fragmentation.

1. Introduction

2. VM Placement Challenges in GPU-Enabled Environments

3. Advanced vGPU Consolidation Feature

4. Case Study 1: 2 Host Servers with 2 GPUs Each

5. DRS Optimization for vGPU Consolidation

6. Case Study 2: 3 Hosts with 4 GPUs Each

7. Benefits of GPU VM Consolidation in vSphere 8 Update 2

Key Sections & Summaries

1. Introduction

• • • Virtual GPUs (vGPUs) are essential for machine learning (ML) and high-performance computing (HPC) workloads.
    • Large models like LLMs (Large Language Models) demand multiple GPUs, requiring optimized resource placement.
    • vSphere 8 Update 2 enhances VM placement strategies to improve GPU utilization.

2. VM Placement Challenges in GPU-Enabled Environments

• • • Problem: Misalignment of VM placement across GPU-enabled hosts leads to inefficient GPU usage.
    • Example: A VM requiring two GPUs may not find a host with two available GPUs due to earlier dispersed placements.
    • This leads to extra administrative work such as vMotion to reallocate resources.

3. Advanced vGPU Consolidation Feature

• • • Introduces a new feature to group similar-sized vGPU profile VMs onto the same host.
    • Admins can enable this feature using the vSphere Client:
      VgpuVmConsolidation = 1
    • This ensures better bin-packing and improves GPU resource utilization.

4. Case Study 1: 2 Host Servers with 2 GPUs Each

• • • Before Update 2:
      • Two single-vGPU VMs are randomly assigned to separate hosts.
      • A new 2-vGPU VM cannot be placed due to fragmentation.
    • After Update 2 (with VgpuVmConsolidation enabled):
      • Single vGPU VMs are grouped onto one host.
      • The 2-vGPU VM can now be placed on the second host.

5. DRS Optimization for vGPU Consolidation

• • • vSphere Distributed Resource Scheduler (DRS) takes GPU profiles into account.
    • Multiple DRS passes may be required to optimize VM placements.
    • To minimize excessive vMotion operations, set:
      LBMaxVmotionPerHost = 1

6. Case Study 2: 3 Hosts with 4 GPUs Each

• • • Before Update 2:
      • Smaller VMs are spread across hosts, preventing a 4-vGPU VM from finding an available host.
    • After Update 2:
      • vGPU consolidation groups VMs, allowing the 4-vGPU VM to be placed correctly.

7. Benefits of GPU VM Consolidation in vSphere 8 Update 2

• • • Efficient GPU resource utilization by avoiding fragmentation.
    • Reduced administrative overhead by eliminating unnecessary vMotion operations.
    • Improved performance for GPU-intensive workloads.

If you want to deep dive into more details, please visit https://blogs.vmware.com/