membench(1)

✨ membench(1) — the memory bandwidth & latency glow-up ✨

okay bestie, welcome to the General Commands Manual section for membench, and lemme tell you, this little icon is everything 💖 no preamble, we're going straight in.

---

🌸 NAME 🌸

membench — modern memory bandwidth and latency benchmarks 🔥

---

🎀 SYNOPSIS 🎀

membench [options]

---

💅 DESCRIPTION 💅

This manual page documents briefly the membench command. (Yes, "briefly" — she's mysterious, she's that girl. 🙈)

membench is a portable, multi-platform memory bandwidth benchmark built for comprehensive system analysis, periodt. 💅

Latency measurement uses linked list traversal with random node order to ✨defeat prefetchers✨ — like, the hardware tries to be sneaky and guess where you're going next and this benchmark is just "no. nope. random." 😭

Statistical validity is ensured via multiple samples until the coefficient of variation is under 5%, or the max sample count is hit — whichever comes first, bestie, we're not torturing your CPU forever.

Memory model: each thread gets its own buffer. everyone minding their own business, as they should 🧚

Total memory = size_kb × threads (×2 for copy, because copy needs src + dst — the girlies need both 💜)

---

🌈 OPTIONS 🌈

• -h — show this help. short, sweet, slay.
• -V — print version and exit. bye babes 👋
• -v — verbose output (use -vv if you want more detail, because of course there's a double version 😌)
• -s SIZE_KB — test only this buffer size, measured in KB. like, e.g. -s 1024 for 1MB. one size only, we're being decisive today 💅
• -f — full sweep: test literally all the sizes up to the memory limit. Default is to test up to 512 MB per thread.
• -p THREADS — use exactly this many threads (default: num_cpus)
• -a — auto-scaling: try different thread counts to find the best one. She's slower but she finds the optimal thread count per buffer size. highkey worth it 👑
• -t SECONDS — maximum runtime, where 0 = unlimited (default: unlimited). no curfew, no cap 🕐
• -r TRIES — repeat each test N times and report the best (default: 3)
• -o OP — run only this operation: read, write, copy, or latency. Can be specified multiple times (default: all). pick your fighter, or pick all of them, we don't judge 🥰
• -H — enable huge pages for large buffers (>= 4MB). Uses THP (no setup needed) or explicit 2MB pages. Automatically skipped for small buffers, because petty for no reason isn't the vibe 💀
• -R — human-readable output with summary (default: CSV). for when you want her in her soft, readable era 📖

---

🦋 ENVIRONMENT VARIABLES 🦋

OpenMP Thread Affinity (environment variables):

• OMP_PROC_BIND=spread — spread threads across NUMA nodes (this is the default)
• OMP_PLACES=cores — one thread per physical core
• OMP_NUM_THREADS=N — override the thread count

---

🍓 CSV OUTPUT 🍓

Output is CSV to stdout, and the columns are:

• size_kb — per-thread buffer size (KB)
• operation — read, write, copy, or latency
• bandwidth_mb_s — aggregate bandwidth in MB/s (0 for latency)
• latency_ns — median memory latency in ns (0 for bandwidth)
• latency_stddev_ns — latency standard deviation in ns (0 for bandwidth)
• latency_samples — number of samples for latency measurement
• threads — thread count used
• iterations — iterations performed
• elapsed_s — elapsed time in seconds

---

💜 SEE ALSO 💜

tinymembench(1), memtester(8).

---

🫶 a lil heads-up from your girl: this page is from the General Commands Manual, dated February 9 2026, and the man page itself is labeled membench(1). So when you cite her, cite her correctly. iykyk 💖

k bye bestie, go benchmark something and let your RAM be the main character ✨🌈💅