Modern cloud-native workloads—especially AI, ML, and HPC—are redefining infrastructure requirements. Traditional virtualization introduces overhead that limits performance, making many organizations rely on bare metal or proprietary platforms.

This session challenges that assumption.

We present a production-grade OpenStack deployment from Vietnam, where infrastructure is optimized to eliminate bottlenecks across:

Compute optimization
CPU pinning, NUMA alignment, HugePages
Network acceleration
SR-IOV, DPDK for near line-rate throughput
Storage performance
NVMe local ephemeral + Ceph for scalable data layers
GPU utilization
PCI passthrough, vGPU, MIG for AI workloads

These optimizations enabled:

Near bare-metal performance for VM workloads
Improved GPU efficiency (~7%) vs proprietary platforms
Support for large-scale AI systems (thousands of GPUs)
High-throughput, low-latency workloads (e.g., real-time AI services)