metadata
language:
- en
license: mit
tags:
- mcp
- benchmark
- gradio
- fastmcp
- api-performance
pretty_name: MCP Server Benchmark Results
size_categories:
- n<1K
🔬 Gradio vs FastMCP Benchmark Report
Generated: 2026-03-04T10:26:47.242725
Total scenarios: 12
Executive Summary
- echo: Gradio wins (151.4 vs 126.2 RPS, 1.2x difference)
- Gradio best config: concurrency_limit=1.0
- async_sleep: Gradio wins (126.6 vs 43.9 RPS, 2.88x difference)
- Gradio best config: concurrency_limit=1.0
Throughput (Requests/Second)
| ('fastmcp', 'mcp_streamable') | ('gradio', 'mcp_streamable') | |
|---|---|---|
| ('async_sleep', 1) | 17.54 | 16.57 |
| ('async_sleep', 10) | 43.87 | 126.56 |
| ('echo', 1) | 126.23 | 54.28 |
| ('echo', 10) | 75.02 | 151.45 |
Latency p50 (ms)
| ('fastmcp', 'mcp_streamable') | ('gradio', 'mcp_streamable') | |
|---|---|---|
| ('async_sleep', 1) | 62.46 | 62.939 |
| ('async_sleep', 10) | 146.304 | 78.523 |
| ('echo', 1) | 7.506 | 0 |
| ('echo', 10) | 115.045 | 62.953 |
Gradio concurrency_limit Scaling
How does Gradio's throughput change as concurrency_limit increases?
| 1.0 | |
|---|---|
| ('async_sleep', 1) | 8.425 |
| ('async_sleep', 10) | 64.665 |
| ('echo', 1) | 27.14 |
| ('echo', 10) | 77.105 |
Protocol Overhead: HTTP API vs MCP
Comparing latency of the same tool called via REST API vs MCP protocol:
| mcp_streamable | |
|---|---|
| ('fastmcp', 'async_sleep') | 104.382 |
| ('fastmcp', 'echo') | 61.2755 |
| ('gradio', 'async_sleep') | 2104.85 |
| ('gradio', 'echo') | 1050.52 |
Error Rates
All scenarios completed with 0% error rate. ✅
Resource Usage
| server | ('mean', 'avg_cpu_pct') | ('mean', 'avg_memory_mb') | ('mean', 'peak_memory_mb') | ('max', 'avg_cpu_pct') | ('max', 'avg_memory_mb') | ('max', 'peak_memory_mb') |
|---|---|---|---|---|---|---|
| fastmcp | 0 | 5.655 | 5.66 | 0 | 5.67 | 5.67 |
| gradio | 0 | 4.955 | 4.955 | 0 | 5.68 | 5.68 |
Benchmark Charts
Methodology
- Both servers use identical tool implementations (imported from shared_tools.py)
- Each scenario runs in an isolated server subprocess
- Warmup period excluded from measurements
- Load generated by async httpx workers (not external tools)
- MCP tests use full protocol lifecycle (initialize → call_tool)
- System metrics sampled every 1s via psutil
Benchmarks generated by mcp-server-bench