MCP Server Performance Benchmark v2
A comprehensive experimental analysis comparing 15 Model Context Protocol (MCP) server implementations across Rust, Java (Spring MVC, WebFlux, Virtual Threads), Quarkus, Micronaut (JVM and GraalVM native images), Go, Bun, Node.js, and Python. Three independent runs totaling 39.9 million requests with I/O-bound workloads (Redis + HTTP API), 0% error rate across all servers.
Results & Analysis
Full results, implementation details, and recommendations:
https://www.tmdevlab.com/mcp-server-performance-benchmark-v2.html
Key Findings
| Server | RPS | Avg Latency | RAM avg |
|---|---|---|---|
| rust | 4,845 | 5.09 ms | 10.9 MB |
| quarkus | 4,739 | 4.04 ms | 194 MB |
| go | 3,616 | 6.87 ms | 23.9 MB |
| java | 3,540 | 6.13 ms | 368 MB |
| java-vt | 3,482 | 9.03 ms | 350 MB |
| quarkus-native | 3,449 | 10.36 ms | 36 MB |
| micronaut | 3,382 | 9.75 ms | 216 MB |
| java-webflux | 3,032 | 8.89 ms | 485 MB |
| java-vt-native | 2,447 | 19.06 ms | 194 MB |
| java-webflux-native | 2,413 | 14.43 ms | 351 MB |
| java-native | 2,316 | 16.20 ms | 178 MB |
| micronaut-native | 2,161 | 20.75 ms | 63 MB |
| bun | 876 | 48.46 ms | 549 MB |
| nodejs | 423 | 123.50 ms | 395 MB |
| python | 259 | 251.62 ms | 259 MB |
All 39.9 million requests completed with 0% errors across all 15 servers and 3 runs.
Project Structure
benchmark-mcp-servers-v2/
├── rust-server/ # rmcp 0.16 + Tokio + deadpool-redis
├── go-server/ # mcp-go + tuned http.Transport
├── java-server/ # Spring Boot 4 MVC (blocking thread pool)
├── java-vt-server/ # Spring Boot 4 + Project Loom Virtual Threads
├── java-webflux-server/ # Spring Boot 4 WebFlux (Reactor / Netty)
├── quarkus-server/ # Quarkus 3.31.4 (Vert.x / Mutiny)
├── micronaut-server/ # Micronaut 4.10.8 / MCP SDK 0.0.19
├── nodejs-server/ # MCP SDK + Express, WEB_CONCURRENCY=4
├── python-server/ # FastMCP + uvloop, 4 workers
├── api-service/ # Go HTTP API (100k in-memory products)
├── infra/ # Redis configuration
├── benchmark/
│ ├── benchmark.js # k6 load test (50 VUs, 5 min, 60s warmup)
│ ├── run_benchmark.sh # Full benchmark orchestration script
│ ├── collect_stats.py # Docker stats collection
│ ├── consolidate.py # Results aggregation across runs
│ └── results/ # Raw results per run (summary.json, k6.json)
├── config/ # Shared benchmark configuration
├── build-server-images.sh # Build all server Docker images
├── docker-compose.yml
└── TMDevLab/ # Published post (HTML/CSS/JS)
Native image variants (Quarkus, Java MVC, Java VT, WebFlux, Micronaut) are built from separate Dockerfiles inside each server directory (Dockerfile.native).
Benchmark Workload
Each server implements three tools backed by real I/O:
| Tool | Operations |
|---|---|
| search_products | HTTP GET /products/search + Redis ZRANGE (parallel) |
| get_user_cart | Redis HGETALL + HTTP GET /products/{id} + Redis LRANGE (parallel) |
| checkout | HTTP POST /cart/calculate + Redis pipeline INCR+RPUSH+ZADD (parallel) |
Test configuration: 50 VUs, 5-minute sustained load, 60-second warmup excluded, Redis flushed and re-seeded before each server test, 2 vCPUs / 2 GB per container.
Running the Benchmark
Prerequisites
- Docker and Docker Compose
- Python 3 (for stats collection and consolidation)
- k6 (optional — the orchestration script uses a containerized runner)
Build Images
./build-server-images.sh
Run Full Benchmark
cd benchmark
./run_benchmark.sh
Results are saved to benchmark/results/<timestamp>/ with a summary.json per run.
Run a Single Server Manually
# Start infrastructure
docker-compose up -d mcp-redis mcp-api-service
# Star
... [View full README on GitHub](https://github.com/thiagomendes/benchmark-mcp-servers-v2#readme)
