LMbench

LMbench is a free, self-hosted command-line benchmark suite for measuring low-level system performance. It suits engineers, administrators and developers who need detailed data on memory, processor, storage and network behavior rather than a single consumer score. Its broad Unix coverage and built-in result comparisons are valuable, while partial Windows support and sensitivity to background activity demand disciplined runs.

Overview

LMbench focuses on data movement and the cost of moving it. It reports latency as elapsed time and CPU cycles per operation, and bandwidth as megabytes per second and CPU cycles per megabyte. That makes it useful for finding where a system spends time or processor capacity, not for producing an all-purpose ranking.

The suite measures memory, disks, file systems, networks, pipes and virtual-memory systems. Process tests include null-process execution time and context switching. The project is GPL-2.0 licensed and deployed on your own machine, with no hosted service or subscription layer. The repository identifies the lmbench 2alpha8 net release; its tags page has no releases, so users should expect a repository-centered workflow rather than a conventional packaged release cadence.

Key features

Detailed latency and bandwidth work

LMbench examines cache-line size, effective TLB size, cache hierarchy, operation latency, memory bandwidth and SysV semaphore latency. The stream benchmark uses kernels from John McCalpin's STREAM and STREAM version 2 benchmarks to report memory bandwidth. These measurements help when tuning memory paths, comparing hardware configurations or investigating why an application is limited by a particular operation.

Process, storage and network coverage

Beyond memory, the workload set includes file-system and disk operations, pipes and virtual memory. Network tests report transfers per second, CPU cycles per transfer, socket, bind and close rates, and throughput for large transfers. This breadth makes LMbench a practical baseline tool for system administrators and performance engineers, although interpreting each result requires more technical judgment than reading a single headline number.

Automation and repository comparisons

The command-line interface supports scripted or repeatable runs. The make see command displays output against other system results included in the repository, and the documentation describes listing results by operating system. That reference data gives measurements context, but it is most useful when the comparison systems and test conditions are understood rather than treated as a universal leaderboard.

Controlled timing

The benchmp() timing harness measures performance under a fixed load and provides controls for parallel processes, warmup and repetitions. These controls are valuable for structured experiments and regression checks. The README warns that other activity can affect results, especially CPU cache and CPU-cycle effects, and recommends several runs with output comparison. LMbench therefore rewards a quiet, repeatable test environment.

Pricing

LMbench has one plan: lmbench — free. It is free software under GPL-2.0, with no free trial because the software is already free. There are no seat, quota, renewal or retention charges stated for this self-hosted plan. Choose it when you can operate the benchmark locally and do not need a vendor-hosted dashboard; readers wanting managed observability or packaged support should consider the alternatives below.

Platforms

LMbench runs on Linux and Windows, with Windows NT support described as partial. Its documentation references UNIX/POSIX systems including Linux, FreeBSD, NetBSD, Solaris, SunOS, AIX, HP-UX, IRIX and OSF/1. The README also gives a LoongArch command using make results OS=mips64el, and mentions Euler and CentOS configuration lines. The repository lists language support for Golang, Java, Python, PHP, .NET, Ruby and Node.js, plus profiling categories covering CPU time, allocation, hardware performance events and hardware metrics.

Who it's for

Pick LMbench for low-level investigations, hardware or kernel comparisons, capacity baselines and automated performance checks on Unix-like systems. It is especially appropriate for teams comfortable with a CLI, self-hosting and interpreting separate latency, bandwidth and operation results. It is a weaker choice for Windows-first environments because support is partial, and for nontechnical users who want a hosted interface, continuous collection or a simplified score.

Pros and cons

• Pro: Broad coverage links memory, cache, TLB, processor, storage, process, file-system and network behavior in one suite.
• Pro: CPU-cycle and per-operation reporting exposes costs that a single throughput score can hide.
• Pro: CLI commands, repetition controls and make see support scripted runs and repository-based comparisons.
• Con: Results can shift with other system activity, particularly cache and CPU-cycle effects, so careless runs can mislead.
• Con: Windows NT support is only partial, reducing confidence on Windows-focused estates.
• Con: No tagged releases means users must manage a repository-based build and update process.

Alternatives

For a directory of ongoing profiling products rather than a standalone benchmark, see Continuous Profiling Tools. For broader desktop-oriented benchmarking choices, see PC Benchmark Software.

• Guance Container Monitoring is a freemium, hosted-and-self-hosted monitoring option with Android, iOS, Linux, macOS, Windows and web access, plus a free tier for up to 10 hosts and daily metric and log limits.
• Middleware suits teams wanting freemium application observability across major desktop, mobile and web platforms, with a free tier capped by data, RUM, synthetic-check and browser-test allowances.
• Grafana Cloud Application Observability is preferable when managed observability is the priority: its free tier includes 2,232 host hours, while Pro adds usage pricing and a monthly platform fee.
• Coroot is a self-hosted monitoring alternative with an Apache 2.0 Community Edition and paid per-monitored-CPU capacity for larger deployments.
• OneUptime fits users seeking an open-source, Apache 2.0 platform that can be inspected, modified and self-hosted.
• Sentry Continuous Profiling is an alternative for continuous profiling across Linux, Windows, macOS, mobile, web and self-hosted environments.
• Metoro Kubernetes Profiling is aimed at Kubernetes profiling, with a free Hobby tier for one cluster, one user and two nodes, plus paid Scale capacity.
• Parca is a free, Apache 2-licensed profiling option for API, Linux, self-hosted and web deployments.

Verdict

Choose LMbench when you need a free, scriptable way to dissect low-level latency, bandwidth and system-operation costs, especially on Unix-like machines. Its breadth and comparison workflow are the main reasons to use it. Look elsewhere if you need dependable Windows coverage, a packaged release stream, continuous profiling or a managed observability service.