The minimalist C server that underpins the 15 million‑instance experiment.
*A lone developer aims to run 15 million web servers on a single VM with 10 GB RAM and one CPU core. The proposal leans on port tricks, massive IP subnets and custom runtimes. Success would, in theory, flatten demand for web hosting and cripple mass‑scale hacking.*
A lone coder set out to prove that scarcity in web hosting is a myth. The goal: spin up 15 million independent web servers inside a single virtual machine armed with just 10 GB of RAM and a single CPU core. The author writes a minimalist C server that listens, accepts, reads, replies "OK" and closes, claiming a raw throughput of about 65 k requests per second. By scaling the process count, hijacking port ranges, and stacking network cards, the experiment promises a world where every human could own a personal web server, slashing the market for shared hosting and, allegedly, ending mass hacking attempts.
The experiment uses port 0 so the kernel picks a free port from ip_local_port_range, which defaults to 32768‑60999. That range yields only about 28 k ports. Ports 1‑1024 need root, 1024‑32768 are skipped to avoid clashes with existing services, and 60999‑65535 are excluded for historic IP masquerading reasons. Even if the range were expanded to the full 1‑65535 space, the total falls far short of 15 million. The author notes that a UDP server can share the same port, doubling the count to roughly 128 k, still nowhere near the target.
The author’s baseline server code listens, accepts, reads, replies with a static "OK" and closes. A single instance reaches about 65 k requests per second. Running 15 million copies would require 15 million processes, each consuming 1.5 MB RAM, totaling roughly 20 TB—far beyond the 10 GB budget. The write‑up suggests collapsing all instances into one process using a custom user‑land scheduler, similar to V8’s multi‑guest model, to cut per‑task memory. No concrete memory figure for that approach is provided, only the claim that it would fit the budget.
To multiply address space, the author proposes attaching 450 network cards, each providing a unique IP address. With 450 IPs and the full 28 k port range, the product reaches the 15 million target. An alternative is AnyIP, which lets a host listen on an entire subnet. IPv6 subnets like 2001:db8:1234:5678::/64 supply 18 quintillion addresses, effectively removing the address limit. For IPv4, acquiring a /24‑size subnet (about 15 million addresses) is estimated to cost $250 million, according to the post.
The post admits that the simple looped server runs out of ports and memory. It also acknowledges that even if the kernel were patched to accept 32‑bit ports, client software would still expect 16‑bit ports, breaking connectivity. The hardware‑card multiplication hinges on a massive, costly NIC deployment. The AnyIP route depends on provider‑granted subnets that are not freely available. While the custom runtime could shrink memory, no benchmark is offered. The overall picture is a thought experiment that highlights limits rather than a viable production plan.
The default ip_local_port_range is 32768‑60999, giving about 28 k ports; ports 1‑1024 need root, 1024‑32768 are skipped to avoid conflicts, and 60999‑65535 are excluded for historic reasons.
Each server consumes 1.5 MB of RAM, so 15 million servers would need roughly 20 TB, far exceeding the 10 GB budget.
The author proposes attaching 450 network cards for 450 unique IPs, using AnyIP to listen on an entire subnet, or leveraging an IPv6 /64 subnet that provides 18 quintillion addresses.
The proposal reads like a provocation more than a production roadmap. It forces a look at the hard limits of TCP ports, memory per process, and the economics of IP address allocation. Until a kernel can truly support billions of sockets and a scheduler can run millions of tasks in a few megabytes, the 15 million‑server VM remains a clever illustration of where software optimism meets hardware reality. The next step will be to test the custom runtime at scale, or to admit that the problem cannot be solved by code alone.
Sources: Hacker News – https://kapteko.com/15-million-webservers/