How close is your server to Baidu AI Cloud?

It measures, from the machine it runs on, the latency to each region of 22 cloud providers: 392 regions in all. For every region you get the median, minimum, P90, jitter (the mean difference between consecutive samples) and loss, and a tier: T0 up to 30 ms, T1 up to 80, T2 up to 150, T3 up to 250, T4 above.

Each region is represented by a public endpoint the provider runs there, such as an API or object storage host, so it tests the network path to that region, not a virtual machine in it.

• TCP handshake (default): the time to open a connection to the region's HTTPS port. One network round trip with no server work in it, the fairest number for comparing regions.
• HTTPS: a request on an open, reused connection, timed until the response headers arrive. It is what an app sees and the same number Me2Cloud shows; some providers add server time, Vultr's looking glass about 100 ms.
• ICMP: a classic ping. Some networks give ICMP lower priority, so it can read higher than TCP, and Naver Cloud does not answer it. It needs root or CAP_NET_RAW.

Latency maps and looking glasses measure from their own datacenters, and Me2Cloud measures from your browser at home. When the question is how a server talks to the clouds, as with an app server and its database, a CDN origin, a backup target or a second provider, the path starts at the server and only a test from there answers it.

The source is on GitHub under the MIT license; every release lists SHA-256 checksums and carries GitHub build provenance, which gh attestation verify checks. The installer downloads the build for your system, checks it and copies it to /usr/local/bin.

Besides the measurement, a run fetches the endpoint list from this site (that request counts runs, by tool version and system only), asks Cloudflare and ident.me which network it runs from, and stores the full share link on s.1ms.dog to make a short one. The result holds per-region numbers and the city and provider, never the IP address. --long-link keeps the result off any server, and --offline makes no request but the measurement.

Each host gets one sample at a time with a 250 ms pause in between, at most four hosts of one provider are measured at once, and a host is given up after three failures in a row. A default run sends each region ten TCP handshakes, the same as opening ten connections to a public website.

A shared link opens here with the server's city and provider, the measurement and the time. A run over IPv4 and IPv6 has a switch between the two in the bar at the top. The tier colors are the same as in the terminal, and the chart places every region by its median.

Links name the endpoint list they were made with, and this site keeps every list it has served, so old links still open after regions are added.

Me2Cloud runs the same kind of test in your browser, from your own line at home or at work. VPS2Cloud runs it on a server. Both use the same endpoint list and the same tiers, so the two results compare directly; "Test from my browser" on a shared result opens Me2Cloud with the same providers.