speed

Internet speed test and live throughput monitor for the terminal.

  speed  Vorboss Limited · City of London (LHR)
  via London · Roubaix · Frankfurt · Nuremberg

  ◷ Latency       4.7 ms   jitter 0.7 ms
  ↓ Download  113.2 Mbps   peak 124.3 Mbps · 139 MB · loaded latency 189 ms (+185 ms)
  ↑ Upload    118.2 Mbps   peak 132.1 Mbps · 150 MB · loaded latency 107 ms (+103 ms)

Install

cargo install --path .

Usage

speed                  # latency, download, upload (10s each)
speed -d 5 --no-upload # test options work without the `test` subcommand
speed test --json      # machine-readable result
speed watch            # continuous download with a live chart; Ctrl+C for a summary
speed watch -s 8 -d 60 # 8 streams, stop after 60s
speed servers          # rank the download mirrors from where you are

For scripts and agents

speed test --json -d 5        # one JSON object (also the default when stdout isn't a terminal)
speed watch --json -d 30      # JSON Lines: {"type":"start"}, {"type":"sample"} per second, {"type":"summary"}
timeout 30 speed watch --json # SIGTERM or Ctrl+C also ends with a summary line

Exit codes: 0 ok, 1 error, 2 bad arguments, 3 a measurement got no data. On 3, the error fields say why (e.g. rate limited (HTTP 429), wait a minute). speed watch --json without -d runs until stopped.

How it measures

• Throughput is counted in-process on its own connections, so other traffic on the machine doesn't skew it. The final rate skips the first 1–2 seconds of ramp-up.
• Download uses the 5 nearest of 34 public mirrors (Hetzner, Linode, OVH, thinkbroadband) across North and South America, Europe, Asia and Oceania, 2 connections each. Servers are ranked by ping (cheap for the mirrors), then the nearest get one HTTP request each to confirm they serve files. speed servers shows the ranking. Upload posts to Cloudflare, which is anycast and so already nearby everywhere.
• Latency is ICMP echo to Cloudflare's nearest edge, over an unprivileged socket (no root needed on macOS; on Linux when net.ipv4.ping_group_range allows it). Otherwise it uses the system ping binary, and failing that, times HTTP requests on a warm connection, marked (http). TCP handshake time isn't used: some network gateways complete handshakes locally, which hides the real round trip. Loaded latency is measured the same way during each transfer, which exposes bufferbloat.
• watch keeps parallel downloads running from the nearest servers and samples the rate once a second. It uses a lot of data: about 3.75 GB per minute at 500 Mbps.

Cloudflare rate-limits repeated uploads (HTTP 429); wait a minute between runs.

License

MIT