Skip to content
kryptonservers

How the filtering works, and how we test it

The stack every node runs, and how we measured each number on this site.

The path of a packet

A packet for a protected service meets four stages, each cheaper than the next would be.

  1. XDP, in the network driver. Before the kernel allocates anything, an eBPF program checks the packet: malformed headers, fragments that reassembly would reject, IPv6 header chains past their limit, traffic to ports nothing listens on, and addresses on the ban list are dropped here. Per-source rate limits for SYN, UDP and ICMP use per-CPU maps, so a random-source flood does not serialise the cores.
  2. The SYN challenge. XDP answers each new TCP connection with a SYN-ACK carrying our own cookie: a keyed hash of the connection, the client's sequence number and the minute. A TC program checks the client's answer and hands the connection to the listening socket. Spoofed floods get an answer nobody reads; the listener only ever sees completed handshakes. The cookie carries window scaling and SACK, so clients without TCP timestamps keep their speed.
  3. The proxy. HAProxy, patched and built by us, terminates TLS and forwards to your origins over kept-alive connections, with health checks, retries to another origin and reloads that keep every connection.
  4. The guard. Our agent reads every request event and scores each client address: request floods, error storms, 404 scanners, user-agent switching, credential stuffing, MAC cracking and account sharing. Past the threshold the address is banned in the proxy and in XDP.

What the node keeps doing on its own

Configuration arrives signed from our control plane. If the control plane is unreachable, the node keeps its last confirmed firewall and configuration, across restarts and reboots. Firewall changes are applied provisionally and roll back on their own if the node cannot reach us afterwards, so a bad rule cannot lock a server out.

Lab results

Measured on our own servers joined by a private VLAN, with the method written up for each run. Floods come from dedicated generators; legitimate clients run alongside and count every failed request.

TestResultHardware
SYN flood, random sources2.0 million packets/s answered in XDP; a real client from 150 addresses: 0 failed requests2x Xeon E5-2690 v4, Intel 82599
ACK flood, random sources1.9 million packets/s dropped in XDP; real client: 0 failed requestssame pair
UDP flood to a verified-only port11.3 million packets/s dropped in the driversame pair
Kept-alive requests, one node653,000 requests/s with every rule on, 500 services, XDP and the hand-offsame pair
New connections, one node100,977 connections/s through the SYN challenge, 0 errorssame pair
XDP cost per packet367 ns on established traffic, 874 ns on new connectionssame pair
L7 attack, 50 sourcesevery attacker banned in 0.24–16 s, 0 of 100 players bannedsame pair
Reload storm25 reloads in 50 s under 530,000 req/s and 4,000 players: 0 errors2x EPYC 9174F, ConnectX-6 Dx
Control plane gonethe node served its last configuration for 6 h 20 min, with a full disk and clock jumps meanwhileE5 node

What these numbers are not

They are per node, on our lab hardware. What a node absorbs in production is bounded by its uplink: no software filter can drop a flood larger than the pipe it arrives on. For floods larger than your own server's uplink, the hosted proxy puts our network in front of yours instead.

Want the raw write-ups?

Ask in the chat; we share the full reports, method and all.