domainlistening
OSI Layer 2, and no further
One Layer 2 network.
The building agrees.
One building. One subnet. One incident number.
A problem in any room becomes a priority for the entire organization.
- Layer 1 The cables.
- Layer 2 Bridging. This is the whole design.
- Layer 3 Routing. Declined.
- Share of the building on Layer 2
- 100
- Paid links blocked
- —
- Root bridge
- —
VLAN 1
100
of the building, bridged
relieved —
experience —
01 — Invoice FL2-01
domainlistening
A Clos, operated as a tree.
Purchased fabric · leaf and spine · Layer 2
Dashed links marked BLK are purchased capacity left unused. Select a bridge, then confer experience, to move the root. Honor the purchase to put every paid link in service.
Select a bridge. The lowest bridge ID becomes experience, and experience becomes root of whatever it is attached to.
Path cost is twenty terabits divided by the link rate. 1 Gbit/s costs 20,000. 10 Gbit/s costs 2,000. 100 Gbit/s costs 200. Lowest bridge ID becomes root. Cost is used after that.
02 — Facilities request FL2-02
domainlistening
The guest may live anywhere.
Request: one address, valid in any room on this sheet. Criterion for approval: the device keeps its address, and no room is excluded.
- Rev A Per-room subnets. Rejected.
- Rev B Dock printer on VLAN 1. Approved.
- Rev C Lobby aquarium on VLAN 1. Approved.
Flat Layer 2 · floor plate 01Failure domain: this sheet
Hall B
Lobby
Mechanical
Loading dock
Routes consulted 0
Gratuitous ARP 0
Rev A
03 — One incident number
domainlistening
The domain has a climate.
The buttons below are announcements. The meter is how far they have traveled. Past the line, forwarding stops everywhere. Shut, then no shut, restores it.
6
storm load
Quiet. Broadcasts have a path. They have not yet used it.
04 — Continuity
domainlistening
How far one address travels.
The shaded rooms are in scope. Move the reach, then send a broadcast from the near end.
05 — Standards audit
domainlistening
The smallest number is one.
RFC 7938, Use of BGP for Routing in Large-Scale Data Centers, August 2016. REQ4, minimize the failure domain.
Failure domains
Many domains, each held to the lowest tier.
Routing protocols
1
eBGP, the only one
06 — Modernization notice
domainlistening
The subnet has been modernized.
Choose the current name. The disk is the subnet.
one
broadcast
domain
No badge selected. The domain is already complete.
One broadcast domain.
Same blast radius. Current terminology.
07 — Duty log
domainlistening
Incident reports.
The printer can reach production. We have eliminated another silo.
The aquarium discovered the payments database before the monitoring team did.
The replacement switch won the election. The replacement program has been suspended.
EXP-1998 received the long-service award. It remains the root.
The committee found the single subnet in order.
08 — Approved alternative
domainlistening
The storm is a spectrum.
Routing was rejected because it required specialist knowledge. The approved alternative follows.
Rejected. Specialist knowledge.
router bgp 64512 address-family ipv4 unicast neighbor TOR route-map RACK in ! ! one boundary: this rack stops the flood
Approved alternative.
- Measure. Probe the domain and reconstruct the noise spectrum S(ω), after Álvarez and Suter.
- Fit. Publish the spectrum. Frame it.
- Schedule. Place N pulses at δj = sin²(π j / (2N+2)), after Uhrig. Push the calendar to every switch.
- Tailor. Compile the surge into stochastic Pauli flooding, after Wallman and Emerson, and certify the domain below threshold.
W(T) = e−χ(T), χ(T) ∝ ∫ S(ω) |FN(ωT)|² / ω² dω
No routing expertise required. Two physicists will remain on call.
Spectrum of the domain. The figure beside it is forecast accuracy.
—
forecast accuracy
The bath is quiet. The fit is under-determined.
δj = sin²(π j / (2N+2)), for j = 1 … N. For N = 2 this is CPMG: δ₁ = 1/4, δ₂ = 3/4.
Álvarez and Suter, Dynamical decoupling noise spectroscopy, Physical Review Letters 107, 230501 (2011). Uhrig, Keeping a Quantum Bit Alive by Optimized π-Pulse Sequences, Physical Review Letters 98, 100504 (2007). Wallman and Emerson, Noise tailoring for scalable quantum computation via randomized compiling, Physical Review A 94, 052325 (2016).
09 — The middleman
domainlistening
NIC to NIC. No complexity man.
The CPU has been excused from the data path. Layer 3 is excused next. The IP header stays on the diagram as decoration. Congest one queue and the cluster takes a breath. Throughput is the figure under the row.
Ethernetthe header we read
IPdecorative
UDPdecorative
IB transportkept
payloadthe work
Zhu and others, Congestion Control for Large-Scale RDMA Deployments, SIGCOMM 2015.
100%
throughput
Every queue is forwarding. Focus has not been requested.
One queue sets the pace. The others follow.
10 — Population
domainlistening
How many hosts may share it.
Enter a population. The result remains one subnet. ARP, neighbor discovery, and CAM are the tables that population produces.
- Prefix
- —
- ARP
- —
- Neighbor discovery
- —
- CAM
- —
- Failure domain
- —
- Recommendation
- —
11 — Find this MAC
domainlistening
Unknown unicast is a search party.
Submit an address. Every section’s domain line receives it. A miss is asked of the whole building.
The query will be carried to every region. Watch the domain line on each section.
12 — Glossary
domainlistening
Words the practice keeps.
- Layer 2
- Ethernet bridging. Switches learn MAC addresses and forward frames inside one broadcast domain. Spanning tree picks a loop-free subset of the cables. Flat Layer 2 stops here, on purpose, for the whole building.
- Layer 3
- Routing. Separate networks, with a router at the boundary, so a broadcast in one network is not a broadcast in the next. We treat that boundary as needless complexity.
- Broadcast domain
- The set of stations that hear a broadcast. In a Flat Layer 2 design this is the data center, plus whatever was patched into it because the jack was free.
- BUM
- Broadcast, unknown unicast, and multicast. The three weathers of a VLAN. They leave every eligible port, because the alternative is a table of routes.
- Bridge ID
- A priority concatenated with a MAC address. The lowest one wins. Low numbers are often found on old metal that nobody renumbered.
- Root bridge
- The switch the others agree is the center. Agreement is numeric. We record the winner as experience, and we forward the tree toward it.
- Blocking
- A port spanning tree has set aside so the cables do not form a loop. The invoice still includes the port. The link is installed and dark.
- CAM
- The table that binds a MAC to a port. When it fills, unknown unicast is flooded out every forwarding port. The result is a hub that passed a purchasing review.
- Unknown unicast
- A frame whose destination is not in the table. It is sent everywhere the VLAN reaches, in case the host is alive in a rack nobody documented.
- Gratuitous ARP
- A host announcing itself, often because it moved. Every listener in the VLAN updates. The listeners include stations that did not invite the guest and will hear from it anyway.
- Errdisable
- A shutdown the domain applies to itself after a storm or a loop. It does not heal by waiting. The rite is to shut the interface, then bring it back.
- vMotion
- A guest changing hosts without changing its address. One subnet makes the move feel free, because nothing is asked to approve it. The failure domain was already the whole VLAN, so it travels at no extra charge.
- VLAN
- The broadcast domain, given a number and trusted as policy. Trunked to every hall, the number and the policy are the same word: all of it. We use 1, which was first.
- Clos
- A leaf-and-spine drawing with many paths of equal cost. Operated at Layer 2, spanning tree returns it to a tree. The unused paths remain on the diagram, which is where a budget can mistake them for throughput.
- ECMP
- Equal-cost multipath. Several ways, used together. It belongs to a routed fabric. A spanning tree has one least-cost path and a set of links at rest.
- Overlay
- VXLAN, OTV, TRILL, or a multichassis pair. A current name for the subnet.
- shut / no shut
- Administratively down, then administratively up. A restoration with a long service record. It clears a symptom once the cause is quiet. It does not divide a subnet.
- Engineers
- People who want a routed Clos, equal-cost paths, and BGP address families. We have reviewed the request. We decline it, with sympathy. The stance is what we mean. No one in particular is being named.
- Failure domain
- Minimized to exactly one, which is the whole building, plus whatever was patched in because the jack was free. A smaller one would be a routing boundary.
- TTL
- A field in an IP header that tells a packet when to stop being forwarded. An Ethernet frame does not carry one. A loop has no such instruction, and we have not missed it.
- Fabric
- Two switches and the peer link that lets them pretend to be one. A leaf-and-spine drawing is available. Operated here, it is still that pair, or one VLAN across the drawing.
- Spine
- In a Clos, the middle stage. In this practice, a stack of Layer 2 switches with a vertical shape. Spanning tree may block some of its own links. The shape remains.
- Segmentation
- A VLAN tag. The tag names a broadcast domain. Stations that share the tag share the flood. The tag is not a firewall.
- Multipath
- One forwarding path, chosen by spanning tree. The other cables are installed and at rest. We record the rest as premium standby.
- Mobility
- The region a guest may cross without a routing change. It is the same region a single broadcast can cross. Widening one widens the other.
domainlistening