A UDP reflection/amplification attack is a Layer 4 volumetric vector, the one behind the record floods of the 2010s from Spamhaus in 2013 to GitHub's 1.35 Tbps in 2018, and one of the classes a DDoS test has to reproduce at scale. The attacker sends small UDP requests carrying the victim's forged address to public services that answer with far larger packets, so the victim is buried in replies to questions it never asked. Each half has its own page: reflection for the forged return address, amplification for the size gap.
Why UDP in particular
TCP makes a source prove itself before it receives anything large. A server answers a SYN with one small SYN-ACK and sends no real data until the handshake completes, which a forged address never does. UDP has no such step. A request arrives, a reply leaves, and the service never learns whether the address on the request was real. The classic reflector protocols (DNS, NTP, SNMP, SSDP, CLDAP, chargen, memcached) all answer without any handshake.
The scale was catalogued in 2014. Christian Rossow's "Amplification Hell" paper measured 14 UDP protocols, and US-CERT advisory TA14-017A published the bandwidth amplification table that is still quoted: SNMPv2 at 6.3, DNS at 28 to 54, SSDP at 30.8, chargen at 358.8, NTP at 556.9. Memcached was added in 2018, at 10,000 to 51,000.
Newer UDP protocols were designed with that table in view. DTLS answers a first hello with a small stateless cookie and starts the large certificate exchange only once the client echoes it. QUIC caps what a server may send to an unverified address at three times what it received. The amplifiers still in use are the older protocols nobody has switched off.
Why it matters under attack
Amplified replies are large packets, so the flood is heavy in bits and comparatively light in packets. The uplink fills long before any router runs short of forwarding rate. That puts the decision upstream: by the time the flood reaches the victim's own edge, the link it arrived on is already full, so filtering has to happen at the provider or the scrubbing tier.
The cheap filter is the source port. Reflected traffic arrives from the service's well-known port (53, 123, 1900, 11211 and so on) aimed at a random high port on the victim, and a network that runs no client for those services can drop inbound UDP from them outright.
Fragmentation weakens that filter. A reply larger than the path MTU is split, and only the first fragment carries the UDP header with its port number. The others carry an IP header and payload and nothing a port rule can match, so a port-only ACL stops the first piece of each reply and passes the rest. A companion rule for non-initial UDP fragments toward the targeted addresses closes the gap, and for most services, which legitimately receive almost no fragmented UDP, it costs nothing.
What a DDoS test measures
The test shapes its traffic the way reflected floods actually arrive: well-known source ports, realistic reply sizes, and a share of fragments. That last part matters because a rule that only sees first fragments looks flawless against unfragmented test traffic.
It then measures where the volume stops (at the provider, the scrubbing tier or the edge router), the bits per second each stage absorbs before goodput falls, and how long reflected traffic reaches the victim before mitigation takes hold. It also confirms that the source-port and fragment rules sit on a device before the narrowest link, not behind it. Where UDP reflection sits among the other volumetric vectors is laid out in Understanding DDoS Attack Vectors.