Stacking, virtual chassis and MLAG

Two or more switches can be made to behave as one: a stack or virtual chassis shares a control plane over dedicated links, while MLAG (multi-chassis LAG) keeps two independent control planes and only makes the downstream device think it is connected to one. Both change what the ports are called, which ports are consumed, which cables or optics are allowed in them and how a failed link behaves. This page lists the technologies per vendor, their cabling and optics rules, the alternatives, and what to watch on the transceiver side.

Stack ring with one control plane versus an MLAG pair with peer link and a dual-homed server

The three approaches

Approach	Control plane	Data plane	Typical scale	Failure domain
Stack (ring/chain of fixed switches)	one (master + members)	shared backplane over stack cables	2–10 units, one rack/closet	whole stack on master failure or software bug
Virtual chassis / cluster (two or more chassis or fixed switches over front ports)	one	forwarding across inter-chassis links	2 (sometimes more), can span rooms	both
MLAG / vPC / VLT / VSX	two independent	peer link + keepalive; each switch forwards locally	2	one switch; the other keeps forwarding
L3 leaf–spine with ECMP (no pairing)	independent	routed	any	one link or box; others unaffected

Technologies by vendor

Vendor	Stacking	Chassis-style virtualisation	MLAG-type pairing
Cisco	StackWise (Catalyst 9200/9300: StackWise-160/480/1T) over dedicated rear cables; FlexStack (2960-X)	StackWise Virtual (Catalyst 9500/9600) over front 40G/100G optics; Nexus 2000 FEX	vPC (Nexus)
Juniper	Virtual Chassis (EX2300…EX4650, QFX5100/5120) over dedicated VCP or converted QSFP+/QSFP28 uplinks	Virtual Chassis Fabric (spine-leaf VC)	MC-LAG; EVPN-multihoming preferred now
HPE Aruba	VSF (AOS-CX 6200/6300) over front ports; backplane stacking on 6300 with cables	VSX (8320/8325/8400) over ISL	VSX is MLAG-style (two control planes)
HPE/H3C	IRF (Comware) over front 10G/40G/100G ports	IRF spans chassis	M-LAG (Comware 7)
Huawei	iStack (S-series) over stack cards or front ports	CSS (cluster switch system, chassis)	M-LAG (CloudEngine)
Dell	stacking (OS10 fixed) over front ports	—	VLT (Virtual Link Trunking)
Arista	none	—	MLAG (peer link over front ports)
Extreme	SummitStack over rear or front ports (EXOS); VOSS clustering	—	MLT/SMLT, vIST (VOSS)
Ruckus	stacking over 10G/40G front ports (ICX)	—	—
Nvidia / Cumulus	none	—	MLAG (CLAG) over peer bond
SONiC	none	—	MCLAG (community/vendor implementations)
MikroTik	none	—	MLAG (RouterOS 7, CRS3xx/5xx)
Brocade FC	—	—	fabric, no pairing needed (SAN design)

Cabling and optics for stacking

Stack type	Cable / module	Optics rules
Dedicated rear stack ports (StackWise, 6300 backplane, iStack cards)	proprietary cables, 0.5–3 m, 40–480 Gb/s aggregate, special connectors	no transceivers; cables are vendor-specific and checked
Front-port stacking / VCP over QSFP+/QSFP28 (Juniper VC, IRF, VSF, iStack over uplinks, ICX)	standard 40G/100G DAC, AOC or optics — the port is converted to stack mode (`request virtual-chassis vc-port set pic-slot 0 port 48`)	the same optics policy as for uplinks applies; DAC is the usual choice within a rack, AOC/SR4/LR4 between rooms (Juniper VC allows long-reach VCP across buildings); converting a port to VCP removes it from the uplink pool
StackWise Virtual / CSS / VSX ISL	2–4 × 40G/100G front ports with standard optics	link members should be identical modules; ISL loss splits the pair
MLAG peer link	2+ × 40/100/400G between the pair (DAC in one rack)	must be the fastest, most redundant link in the pair; peer keepalive over OOB

Stack cables have EEPROMs of their own on several platforms; a non-genuine or damaged cable shows as a member that never joins.

What stacking changes for the operator

Aspect	Effect
Port names	first number becomes the member ID: `Gi2/0/1`, `xe-1/0/48`, `2:49` (Port naming & LEDs)
Uplinks	spread across members (one per member) so a member failure keeps the stack connected; LAG across members
Optics inventory	`show inventory` / `show chassis hardware` lists modules per member — the same PN may be accepted on one member's software image and not on a mixed-version stack
Software upgrade	whole stack (ISSU on some); a stack reload takes every optic down
Master election	on failure, ports keep forwarding (NSF) but management moves
Ring vs chain	ring survives one cable failure; chain does not
Mixed models	limited by vendor (same family, same license, sometimes same model)

MLAG specifics

Item	Rule
Peer link	2 × the largest uplink speed, LAG; carries orphan-port and failover traffic
Keepalive	separate path (OOB) to detect split-brain
Downstream	server or switch with an LACP LAG whose members go to both peers — identical speed/FEC on both members (Forwarding basics)
Consistency	VLANs, MTU, STP settings and FEC on member ports must match on both peers; mismatch → LAG member suspended
Optics	a downstream host's two DACs land on two switches: buy and code them as a pair
Modern alternative	EVPN multihoming (ESI-LAG) removes the peer link; pure L3 to the host removes pairing altogether

Symptom	Cause	Where to look
Stack member never joins, stack LED off	stack cable seated wrong, not genuine, or port not converted to VCP	cable EEPROM/LED, `show switch stack-ports`, `show virtual-chassis vc-port`
VCP link flaps, members re-elect	marginal optics/DAC on the VCP	DDM on the VCP port (Link flapping)
MLAG member suspended on one peer	FEC/speed/MTU mismatch between the peers' member ports	Port configuration recipes
Half the servers lose a path after upgrade	one peer's new release rejects third-party optics	Compatibility matrices & firmware
Peer link down, split-brain	peer-link DAC failure with keepalive misconfigured	design: two members on different cards

In CodingBox

Stacking and MLAG double the places a module must be accepted: the same coded identity has to pass on both members or both peers, sometimes on two software versions. Read and record the modules of a pair together, and treat VCP-converted uplinks as uplinks in the inventory (Check transceiver, Code database).