Can SmartX ECP support mission-critical systems like Oracle and GaussDB databases?

Deep Analysis

The concern that HCI “can’t match SAN storage performance” is rooted in early HCI products. SmartX ECP’s optimizations have largely closed the performance gap.

Why Databases on HCI Were Historically Challenged

Early HCI storage architectures created database performance concerns:

• Network Storage Bottleneck: All I/O traversed the network to shared storage
• CPU Overhead: Storage virtualization consumed CPU cycles
• No RDMA Support: Traditional TCP/IP storage networking added latency
• Memory Access Patterns: Cross-NUMA-node memory access degraded performance

SmartX ECP’s Performance Optimizations

1. RDMA (Remote Direct Memory Access)

RDMA enables direct memory-to-memory data transfers without CPU involvement:

• Ultra-Low Latency: Bypasses TCP/IP stack — single-digit microsecond latency
• High Throughput: 100Gbps+ bandwidth with minimal CPU overhead
• Reduced CPU Usage: Storage networking doesn’t compete with application workloads
• Critical for Analytics: Essential for large-scale data processing and real-time analytics

2. Boost Mode (vHost Acceleration)

Kernel-space I/O processing eliminates user-space overhead:

• Path Shortening: VM I/O requests handled in kernel space vs. user space
• System Overhead Reduction: Fewer context switches and memory copies
• Higher IOPS: Sustained higher IOPS output under load
• Lower Latency: Faster I/O completion for transactional workloads

3. Persistent Cache

Persistent cache accelerates access to frequently accessed data:

• Hot Data Acceleration: Frequently accessed data served from persistent cache
• Database Response Improvement: Reduces disk I/O for repeated queries
• Write Acceleration: Write coalescing reduces write amplification

4. CPU Dedication and NUMA Optimization

Critical VMs receive guaranteed CPU and memory resources:

• CPU Dedication: Physical CPU cores exclusively assigned to specific VMs
• No Resource Contention: Critical databases never compete for CPU
• NUMA Awareness: VM memory allocated on the same NUMA node as its vCPUs
• Cross-NUMA Avoidance: Eliminates the latency penalty of accessing remote memory

5. SR-IOV Network Passthrough

For the most demanding network workloads:

• Physical NIC Partitioning: SR-IOV creates multiple virtual functions from one physical NIC
• Direct VM Access: VM accesses physical NIC without hypervisor intervention
• Maximum Throughput: Near-native network performance for critical workloads
• Latency Reduction: Eliminates virtual switch overhead for latency-sensitive traffic

Database Workload Validation

| Database | Key Requirements | SmartX ECP Optimization |
|———-|—————–|————————|
| Oracle OLTP | High IOPS, low latency, strong consistency | RDMA + Boost mode + NUMA + CPU dedication |
| GaussDB | High concurrency, distributed architecture | Persistent cache + NUMA + SR-IOV |
| SQL Server | Consistent latency, strong transaction support | NUMA + CPU dedication + Boost mode |
| PostgreSQL | Open-source, flexible, high availability | NUMA + persistent cache |

Real-World Performance Results

SmartX ECP has been deployed for production database workloads:

• Financial Services: Banks and securities companies running core trading databases
• Manufacturing: ERP and MES systems handling production-critical transactions
• Healthcare: Hospital information systems with 24/7 availability requirements
• Logistics: Real-time tracking and dispatch systems