15 Million Web Servers on One VM: Kernel Bypass With smoltcp, DPDK and AF_XDP
Artem Yarulin runs 15 million HTTP servers on a single VM with 10 GB RAM and one CPU core by replacing the kernel network stack with userland alternatives.

By Ryan Chua, Cloud & Enterprise Tech
, updated 4 min read
Artem Yarulin, a software engineer based in Finland, has demonstrated running 15 million concurrent HTTP servers on a single virtual machine with just 10 GB of RAM and one CPU core. The experiment, detailed on his blog kapteko.com, replaces the Linux kernel network stack with userland alternatives, smoltcp, DPDK and AF_XDP, to achieve unprecedented server density.
Why the kernel stack limits scale
The Linux kernel network stack allocates roughly 3 KB of memory per TCP connection for socket buffers, socket structures and related metadata. At 15 million connections, that alone would require 45 GB of RAM, far beyond the 10 GB available in the test VM. The kernel also serialises packet processing through interrupt handlers and softirqs, creating CPU contention that becomes a bottleneck long before memory runs out.
Userland networking changes the equation
Yarulin's approach moves the entire TCP/IP stack to user space:
smoltcp provides a lightweight TCP/IP implementation designed for embedded systems, with minimal per-connection state.
DPDK (Data Plane Development Kit) polls network packets directly from the NIC, bypassing kernel interrupts entirely.
AF_XDP (AF eXpress Data Path) enables zero-copy packet transfer between the NIC driver and user space via XDP sockets.
Together these reduce per-connection memory to approximately 200 bytes, a 15x improvement over the kernel stack. The single CPU core handles all 15 million connections because DPDK polling eliminates interrupt overhead and smoltcp's event loop processes packets in tight batches.
The test setup
The VM runs on a single core with 10 GB RAM. A custom Rust application uses smoltcp for TCP state machines, DPDK for packet I/O via AF_XDP, and a minimal HTTP parser. Each server responds with a static "Hello, World!" payload. The benchmark establishes all 15 million connections, then measures sustained request throughput.
What this means for infrastructure
While 15 million servers on one VM is a synthetic benchmark, the underlying principle applies to real workloads: kernel bypass networking can increase connection density by an order of magnitude. This matters for edge computing, high-frequency trading, and any scenario where millions of idle or low-rate connections must be maintained cost-effectively.
Cloud providers already offer DPDK-enabled instances (AWS ENA Express, Google Cloud NCCL, Azure Accelerated Networking). The smoltcp crate is available on crates.io for Rust projects. AF_XDP requires Linux 4.18+ and a compatible NIC driver.
Caveats
The benchmark uses raw HTTP without TLS, which adds significant per-connection overhead. Real applications need encryption, request routing, and application logic, all of which increase memory and CPU demands. The single-core design also means no horizontal scaling within the VM. Yarulin notes the next step is adding TLS support and exploring multi-core scaling with sharded event loops.
The full article with code snippets and benchmark methodology is at kapteko.com/15-million-webservers.
Sources
- 1kapteko.com — 15 million web servers on one VMkapteko.com
- 2Hacker News discussionnews.ycombinator.com