Loops, Spanning Tree, and Link Aggregation

Redundant layer-2 links are good for survivability and catastrophic for stability at the same time. The moment two switches have more than one path between them, you have a loop — and a broadcast frame in a layer-2 loop circulates forever, multiplying at every switch, until it saturates every link in seconds. This is a broadcast storm, and it takes a network down faster than almost any other failure because nothing at layer 2 stops it on its own.

Spanning Tree Protocol exists to make redundant topologies safe: it computes a loop-free tree and blocks the redundant links, keeping them as hot standbys that unblock only when the active path fails. Link aggregation solves a related problem from the opposite direction — it bonds several physical links into one logical link for bandwidth and failover, with no loop because the switches treat the bundle as a single port. Both deal in redundancy; they are mirror images of each other.

Why Layer-2 Loops Are Catastrophic

The IP header has a TTL that decrements at every router and kills a looping packet after at most 255 hops. The Ethernet header has no such field — there is nothing in a frame to count hops or expire it. A broadcast frame that enters a loop is flooded out every port, returns to where it started, and is flooded again, with no mechanism anywhere in the frame to stop it.

Worse, each switch in the loop duplicates the frame, so traffic grows exponentially rather than just circulating. Within seconds the links are at 100% utilization carrying nothing but copies of looping broadcasts, CPUs are pinned processing them, and the network is effectively dead for real traffic. A single mistakenly cabled redundant link, with no loop protection, is enough to cause this — which is exactly why STP is on by default on managed switches.

Spanning Tree Protocol

STP builds a loop-free tree in three steps. First, the switches elect a root bridge — the switch with the lowest bridge ID (configurable priority plus MAC). Second, every other switch finds its lowest-cost path to the root and marks that port as its root port. Third, for each segment one switch is chosen to forward, and every remaining redundant port is put into a blocking state — it stays up and listens for topology changes but forwards no data, so no loop can form.

When the active path fails, a blocked port can transition back to forwarding and restore connectivity over the redundant link. That is the entire value proposition: physical redundancy that is logically loop-free, with automatic failover. The cost is that half your links may sit idle in blocking state, carrying nothing until something breaks — which is part of why people reach for link aggregation or layer-3 designs instead.

# spanning tree state on a Linux bridge
mstpctl showport br0
# eth1  forwarding  (root port toward the root bridge)
# eth2  blocking    (redundant link held down to break the loop)

RSTP and Faster Convergence

Classic STP (802.1D) is slow. A port moving from blocking to forwarding walks through listening and learning states on timers, taking 30 to 50 seconds to converge after a topology change — an eternity during which traffic over the recovering path is black-holed. On a network where links flap, that pause is repeatedly mistaken for an outage when it is actually STP reconverging.

Rapid Spanning Tree (802.1w, RSTP) replaces the timers with an explicit handshake between neighbors and converges in under a second on point-to-point links. RSTP is the modern default and supersedes plain STP everywhere it can — if you are still seeing 30-second reconvergence pauses, you are running legacy STP and should move to RSTP or a layer-3 design that avoids the problem entirely.

Link Aggregation

Link aggregation (LACP, 802.3ad, or "bonding" on Linux) bundles multiple physical links into one logical link. Two switches joined by four 10 Gbps links present a single 40 Gbps pipe, and the loss of any one member degrades capacity rather than dropping the connection. Because the switches treat the bundle as one logical port, STP sees no loop — the redundancy is used instead of blocked.

The catch is how traffic spreads across members: LACP hashes each flow (typically source/destination IP and port) to one member and pins it there, so a single TCP connection rides one physical link and cannot exceed that link's speed. Four bonded 10 Gbps links give 40 Gbps of aggregate across many flows, but any one flow tops out at 10 Gbps. Expecting a single large transfer to use the whole bundle is the most common LACP misunderstanding.