SockBench • Project

Overview

SockBench is a Linux-based benchmarking tool designed to evaluate and compare different communication paths—RDMA (Remote Direct Memory Access) vs traditional TCP/UDP sockets. It provides reproducible measurements of latency, throughput, and scaling behavior across message sizes and connection modes.

The goal is to simplify the otherwise complex setup of RDMA-capable environments. With a single lightweight toolchain you can experiment with communication primitives, analyze bandwidth utilization, and build intuition about when low-latency RDMA is worth the complexity over conventional sockets.

Built as part of a systems-level exploration into high-performance networking, SockBench emphasizes clarity, minimal dependencies, and extendability (future: UDP, multi-threaded clients, epoll-based server, distributed multi-host runs).

Key Features

RDMA vs TCP/UDP (extensible design; current prototype focused on sockets)
Latency and throughput measurement with automatic CSV logging
Configurable message sizes & number of messages
Supports repeated batch benchmark runs via script
Lightweight CLI interface (no heavy dependencies)
Generated plots (latency & throughput vs message size)

Usage (from README)

Compile
g++ server.cpp -o server
g++ client.cpp -o client
Run
Terminal 1 (server): ./server
Terminal 2 (client): ./client 1024 10000
Batch Benchmarks
./bench.sh (iterates across message sizes; outputs to results.csv)
Plot Results
python3 results/plot_results.py → outputs latency.png, throughput.png

Sample Results

Latency and throughput vary non-linearly with message size; small packets incur per-call overhead while larger packets saturate bandwidth.

Message Size	Avg Latency (ms)	Throughput (MB/s)
32 B	0.0618	0.49
64 B	0.179	0.34
512 B	0.159	3.07
1 KB	0.097	10.0
4 KB	0.053	73.8
64 KB	0.030	2049.9

Latency vs message size plot — Latency decreases as message size increases until system overhead amortizes.

Throughput vs message size plot — Throughput climbs rapidly with larger messages, approaching link saturation.

Challenges & Learnings

Explored Linux networking stack & soft-RoCE (rxe) constraints in virtualized setups
Balanced measurement stability vs runtime (warmup, batching, clock resolution)
Clarified trade-offs: RDMA’s low-latency path vs generality & simplicity of TCP/IP

Roadmap

Add UDP benchmarks for cross-protocol comparison
Introduce multi-threaded / multi-connection client
Non-blocking epoll server for scalability
Distributed multi-host orchestration

License

MIT License.