Stride - Overall Design Goal

[dragonJACson]

Stride - Modern RDMA Benchmark Suite

Stride is a comprehensive RDMA benchmark suite that goes beyond traditional perftest functionality. It provides real-world traffic patterns that simulate actual business workloads, making it easier to evaluate RDMA performance in production-like scenarios.

Why Stride?

Unlike traditional RDMA benchmarks that focus on basic operations, Stride simulates real-world usage patterns from:

• Cloud Block Storage (EBS-like) frontends/backends
• High Performance Computing (HPC)
• Key-Value Caches
• ML/AI Training Systems
• Mixed workload scenarios

Components

🚀 stride-run

Advanced performance evaluation tool that extends traditional perftest capabilities:

Performance Features

• Multi-threaded benchmark engine
• NUMA-aware memory allocation
• Automatic thread-to-NUMA binding
• Hardware-accelerated checksum verification
• Mixed traffic pattern support
• Comprehensive vendor detection (NVIDIA, Intel, etc.)

Developer Experience

• Rich TUI (Terminal User Interface)
• Structured output formats (JSON/YAML/TOML)
• Support for both CQ and CQ Ex APIs
• Clean, modular codebase
• Send-with-immediate operations

📊 stride-bench

Real-world workload simulator with predefined traffic patterns:

Traffic Patterns

• EBS-style block storage operations
• HPC communication patterns
• KV cache access patterns
• Mixed operations (Write + Send-with-Imm notifications)
• GPU-aware operations with write fencing

Suite Management

• YAML / TOML-based traffic pattern definitions
• Shareable benchmark configurations
• Reproducible test scenarios
• Real-time monitoring interface
• Historical result comparison

🔍 stride-probe

System topology analyzer focused on RDMA infrastructure:

Discovery Features

• RNIC topology mapping
• NUMA node relationships
• GPU affinity detection
• PCIe configuration analysis
• Link rate verification
• Optimal parameter suggestions