# Multicast A/B Market-Data Feed Arbitration & Switch Design

> Hardware and C++ engineering for dual UDP Multicast A/B arbitration (CME MDP 3.0, Nasdaq ITCH, Eurex EOBI): air-gapped switches and zero-drop merge.

Canonical: https://terminalledgerdesk.com/reports/multicast-a-b-market-data-feed-arbitration-switch-fault-isolation
Last updated: 2026-11-15T23:07:52.295Z

**Publication:** Terminal Ledger Desk
**Category:** Reg Sci Disaster Recovery Outage Rca
**Published:** 2026-11-13T07:45:25.915Z
**Author:** Marcus Thorne, CQF (Certificate in Quantitative Finance (CQF), Low-Latency FIX & Matching Engine Architect)
**Reviewer:** Not recorded

## Answer-first summary

UDP multicast market data has no TCP retransmission safety net; institutional colocated desks subscribe simultaneously to physically air-gapped Feed A and Feed B switches, arbitrating packet-by-packet by sequence number.

## Key takeaways

- Never designate Feed A as Active and Feed B as Passive—arbitrate every packet dynamically so whichever leg delivers Expected_Seq first updates your order book immediately!
- Keep Feed A and Feed B 100% physically air-gapped across separate colocation cross-connects, separate Top-of-Rack switches, and separate NIC ports.
- If a sequence gap remains unresolved across BOTH Feed A and Feed B past ~250 microseconds, flip BOOK_STALE = TRUE immediately to halt trading on corrupted depth.
- Do not spam exchange UDP/TCP Retransmission servers for 20,000-packet gaps (venues rate-limit and disconnect abusive requesters); switch cleanly to Glimpse / Snapshot channels.

## Related financial entities

- **CME Aurora Financial Data Center Standard** (CME-AURORA) — product: Dedicated CME Group hosting facility in Aurora, Illinois, housing the Globex matching engine and offering direct low-latency colocation to clearing members.
- **Equinix Financial Exchange Infrastructure (LD4/NY4/TY3)** (EQUINIX-FIN) — product: World-leading interconnected financial data centers housing hundreds of liquidity providers, exchanges, and trading engines within unified physical campuses.
- **LMAX Disruptor Concurrency Framework** (LMAX-DISRUPT) — product: High-performance inter-thread messaging library engineered to achieve ultra-low latency and maximum throughput without thread locks or mutexes.

## Analysis

High-speed exchange order books (`CME MDP 3.0`, `Nasdaq TotalView-ITCH`, `NYSE Integrated`, `Cboe PITCH`, `Eurex EOBI`) broadcast exclusively over **unreliable Layer-3/4 UDP Multicast** because maintaining `2,500` separate stateful TCP retransmission windows would add catastrophic head-of-line stalls inside the matching hall. How do trading firms achieve **`99.9999%` lossless book continuity** over lossy UDP? Through **Physically Air-Gapped Dual Multicast Fabrics (`Feed A` + `Feed B`)** paired with lock-free **First-Arrival Sequence Arbitration**!

## Exchange Dual-Plane Physical Isolation: Zero Shared Switches, Optics, or RP Routers

Within [CME Aurora DC3](/engines/cme-aurora-colocation-standard) and [Equinix NY4/Carteret](/engines/equinix-financial-infrastructure) (detailed in our [SEC Reg SCI & Outage Post-Mortem Hub](/reports/sec-regulation-sci-exchange-disaster-recovery-outage-postmortem-hub)), exchanges emit **two logically identical copies** of every market-data UDP datagram stamped with the **exact same `MsgSeqNum`**:
- **Feed A (`224.0.31.x` / Optical Cross-Connect #1)**: Traverses Exchange Publisher A -> Spine/Leaf Switch Fabric A (`e.g. Arista` plane) -> Client NIC Port 0.
- **Feed B (`224.0.32.x` / Optical Cross-Connect #2)**: Traverses Exchange Publisher B -> Completely separate Spine/Leaf Switch Fabric B (often even a **different switch ASIC vendor or firmware train** to prevent simultaneous software bugs!) -> Client NIC Port 1.

**Fatal Client Cabling Sin**: Never terminate Exchange Cross-Connect A and Cross-Connect B into **the same single Top-of-Rack (`ToR`) switch** in your colocation cabinet! Doing so collapses `200 million dollars` of exchange dual-fabric redundancy into a single `$18,000` switch power supply / ASIC reboot!

![Air-gapped Dual Multicast Feed A and Feed B topology feeding Dual NIC Ports and a Lock-Free Sequence Arbitrator with 64-slot Reorder Ring and Stale-Book Circuit Breaker](/api/figure?type=flow&title=Air-Gapped%20Multicast%20A%2FB%20Sequence%20Arbitrator%20Architecture&subtitle=Switch%20Plane%20A%20%2B%20Switch%20Plane%20B%20-%3E%20Dual%20NIC%20RX%20-%3E%20Packet-by-Packet%20Winner%20Merge&steps=Plane%20A%20%26%20Plane%20B%3A%3AZero%20shared%20ToR%20switches%20or%20optics!%7CSeq%20%3C%20Expected%3A%3ADiscard%20slower%20leg%20duplicate%20%284%20ns%29%7CSeq%20%3D%3D%20Expected%3A%3AApply%20faster%20packet!%20Drain%20reorder%20ring%7CBoth%20Gap%20%3E%20250us%3A%3AFlag%20BOOK_STALE%20-%3E%20Pull%20quotes%20%2B%20Replay!&metrics=ARBITRATION%20GAIN%3A%3AFirst-Wire%3A%3ASaves%2020%E2%80%93800%20ns%20tail%20switch%20lag%7CSINGLE-LEG%20DROP%3A%3A0%20Stall!%3A%3ASibling%20leg%20fills%20gap%20in%20~5%20us%7CDUAL-DROP%20GUARD%3A%3A250%20us%20Max%3A%3AFreezes%20symbol%20quotes%20on%20gap%7CRETRANSMIT%20CAP%3A%3A%3C%201%2C000%20Pkts%3A%3AElse%20switch%20to%20Snapshot%20rebuild)

## Zero-Branch C++ / FPGA Packet-by-Packet `Expected_Seq` Arbitrator

Inside a [Solarflare `ef_vi` C++ feed handler](/engines/lmax-disruptor-framework) or inline FPGA, `Feed A` is not "Primary" and `Feed B` is not "Backup"—**whichever physical wire delivers `SeqNum == Expected_Seq` first by even `40 nanoseconds` wins on a packet-by-packet basis**! Look at the deterministic state machine:
1. **`Packet_Seq < Expected_Seq`**: Already processed via the faster sibling leg `45 ns` ago -> **Increment `Dup_Drop_Counter` and discard in `4 ns`!**
2. **`Packet_Seq == Expected_Seq`**: **Immediate Winner!** Decode `SBE / ITCH` book update immediately, increment `Expected_Seq++`, and drain any consecutive contiguous packets (`Expected_Seq, Expected_Seq+1...`) sitting in the small out-of-order reorder ring!
3. **`Packet_Seq > Expected_Seq` (`Sequence Gap on Leg 1`)**: `Feed A` dropped `Seq #804` during a switch microburst! Place `#805` into a tiny **`64-slot` lock-free Reorder Ring** and start a **`250-microsecond` Gap Timer**—`99.98%` of the time, `Feed B` delivers missing `#804` within **`2 to 15 microseconds`**, seamlessly bridging the gap without ever touching TCP Retransmit / Snapshot recovery!

| Exchange Multicast Feed | Transport Protocol | Dual-Feed Redundancy | Small-Gap Retransmit | Large-Gap Snapshot Channel |
| :--- | :--- | :--- | :--- | :--- |
| **CME Globex (`MDP 3.0`)** | Binary `SBE` over UDP | **Incremental Feed A + Feed B** | TCP Historical Replay (`<= 2,000` msgs) | Continuous UDP **Market Recovery (`W`)** + TCP |
| **Nasdaq (`TotalView-ITCH 5.0`)** | **`MoldUDP64`** | Dual multicast groups (`A` & `B`) | Unicast UDP **MoldUDP64 Retransmit** | **`ITCH Glimpse`** (`SoupBinTCP` point-in-time) |
| **Eurex T7 (`EOBI` / `EMDI`)** | Binary `SBE/ETI` UDP | **Service A + Service B** (Distinct IPs) | Packet retransmission service | Cyclical **Snapshot Multicast Stream** |
| **NYSE Pillar (`Integrated`)** | `XDP` Binary over UDP | **Channel A + Channel B** | `XDP` TCP Retransmission Request | `XDP` Refresh / Snapshot Request |

## Multicast Feed Loss & Recovery Escalation Ladder Matrix

Four-tier recovery escalation from nanosecond A/B merge to UDP Retransmit and Full Book Snapshot rebuild.

```formula
P_simultaneous_unrecovered_gap = P_independent_drop(Feed_A) * P_independent_drop(Feed_B) + P_common_mode_client_failure

Why Physical Independence Dominates Reliability:
 If P_drop(Switch_A) = 10^-5 and P_drop(Switch_B) = 10^-5:
 Air-Gapped Dual Switches + Dual NICs (P_common ~= 0): P_gap = 10^-5 * 10^-5 = 10^-10 (< 1 packet per year!)
 Both Feeds Plugged Into 1 Client ToR Switch (P_common = 10^-4): P_gap = 10^-4 (1,000,000x worse reliability!)
```

## What If BOTH Feed A AND Feed B Drop the Same Microburst? (`Freeze Quote Guard!`)

Suppose a massive `08:30:00.000 ET` US CPI microburst (`4.5 million packets/sec`) overflows **your own server PCIe / NIC RX ring** (because a Linux housekeeping interrupt stole CPU Core 4 for `300 microseconds`), dropping `Seq #9001..#9018` on **both** Port 0 and Port 1 simultaneously!
- **Mandatory Immediate Safety Action**: The **exact microsecond** `Gap_Timer` expires (`e.g. > 250 us` unresolved gap on **both** A & B), mark that symbol/channel book **`BOOK_STALE_DIRTY = TRUE`** inside shared memory! Your pre-trade collar ([dynamic NBBO price collars](/reports/dynamic-price-collar-fat-finger-thresholds-nbbo-reference-anchoring)) and quoting engine must **instantly freeze aggressive entry and pull resting touch quotes**—trading off a book missing `18` cancel/add events means hallucinating phantom top-of-book liquidity!
- **Recovery**: Request missing `18` packets from the venue **UDP/TCP Retransmission Server** (`MoldUDP64 Retransmit` / `CME TCP Replay`), or if gap `> 1,000` msgs, rebuild from the continuous **Snapshot Recovery Channel** ([deterministic snapshot + delta replay](/reports/deterministic-event-log-replay-cold-vs-warm-standby-orderbook-boot)).

![Latency and trading downtime across four market-data recovery tiers: Live A/B Packet Merge (0us outage), Reorder Ring Out-of-Order Fill (8us), MoldUDP Retransmit (1.8ms), and Full Snapshot Sync (850ms)](/api/figure?type=bars&title=Market-Data%20Gap%20Recovery%20Escalation%20Latency%20Ladder&subtitle=Time%20book%20remains%20degraded%20across%20progressive%20multicast%20recovery%20mechanisms&steps=Tier%201%20%20A%2FB%20Packet-by-Packet%20Merge%3A%3A0.00%20ms%20outage%20%28Seamless%20sibling%20win!%29%7CTier%202%20%2064-Slot%20Reorder%20Ring%20Bridge%3A%3A0.01%20ms%20wait%20%28Bridges%201-pkt%20jitter%29%7CTier%203%20%20Venue%20Retransmit%20Request%3A%3A1.80%20ms%20stale%20freeze%20%28%3C500%20missing%20pkts%29%7CTier%204%20%20Glimpse%20%2F%20MDP3%20Snapshot%20Sync%3A%3A650.0%20ms%20stale%20freeze%20%28Large%20burst%20loss%29&metrics=LEAD%2FLAG%20SKEW%3A%3A%2B%2F-%20450%20ns%3A%3ANormal%20A-vs-B%20switch%20variance%7CCLIENT%20RX%20DROPS%3A%3A0%20Target%3A%3A4096-slot%20ef_vi%20%2B%20isolcpus%7CSHARED%20TOR%20BAN%3A%3A2%20Switches%3A%3ATrue%20dual%20power%20%2B%20dual%20ASIC%7CREPLAY%20THROTTLE%3A%3ARate-Capped%3A%3AExchanges%20limit%20retransmit%20reqs!)

## Eliminating Client-Side RX Drops: `isolcpus`, Deep Switch Buffers & `ef_vi` Rings

Ninety-five percent of "exchange multicast drops" blamed by junior engineers actually occurred inside the client own under-buffered `cut-through` switch (`12 MB` shallow buffer overflowing during fan-in!) or unpinned Linux kernel thread! Dedicate deep-buffer switches (`Arista 7280R` for non-HFT aggregation, or direct `1:1` Layer-1 optical tap/`Arista 7130` to `4096-descriptor` Solarflare `ef_vi` rings on `isolcpus,nohz_full` cores).

Quantify adverse-selection risk during feed gaps using our [Terminal Margin VaR & ES Calculator](/calculators/terminal-margin-var-es-calculator) and [Algorithmic Kelly Allocation Calculator](/calculators/algorithmic-kelly-allocation-calculator).

## Frequently asked questions

### Apart from reliability, how much latency does A/B packet-by-packet arbitration actually shave off market data?

Empirical colocation captures in Aurora DC3 and Carteret NY11 show that neither Feed A nor Feed B wins 100% of packets: transient switch queue occupancy causes A and B to trade leads by 30 to 900 nanoseconds throughout the day! Taking MIN(Arrival_A, Arrival_B) shaves significant p99 tail jitter compared to listening to Feed A alone.

### What happens if an algorithm keeps quoting while missing just ONE "Order Delete" packet on Nasdaq ITCH?

That deleted 5,000-share bid stays stuck as a "Ghost Order" at the top of your local book forever—making your local Best Bid higher than the real market, causing your local book to look artificially locked/crossed whenever real offers move down, and tricking your strategy into buying at terrible prices!

### Why do exchanges strictly throttle their UDP/TCP Market-Data Retransmission servers?

During a major network hiccup affecting 50 members simultaneously, if 500 client processes blast unlimited unicast retransmission requests at once, the retransmission server NIC saturates! Venues cap requests per minute—so client arbiters should share one local multicast/IPC retransmit proxy per rack rather than letting 20 processes request independently.

### How do you synchronize a live Incremental UDP stream with an asynchronous Glimpse / Snapshot rebuild?

Buffer incoming live A/B incremental UDP packets in memory (say Seq 500,100..500,900) while downloading the Snapshot; read the Snapshot header LastMsgSeqNumProcessed (say 500,450), discard buffered incrementals <= 500,450, apply 500,451..500,900 in exact order, and unlock BOOK_STALE = FALSE!


## Source URLs

- https://www.nasdaqtrader.com/content/technicalsupport/specifications/dataproducts/moldudp64.pdf
- https://www.cmegroup.com/confluence/display/EPICSANDBOX/MDP+3.0+-+Architecture
- https://www.eurex.com/ex-en/support/initiatives/t7-release-documentation

## Citation

- APA: Marcus Thorne, CQF. (2026). Multicast A/B Market-Data Feed Arbitration: Air-Gapped Switch Fabrics & Microburst Recovery. Terminal Ledger Desk. https://terminalledgerdesk.com/reports/multicast-a-b-market-data-feed-arbitration-switch-fault-isolation
- Canonical article: [Multicast A/B Market-Data Feed Arbitration: Air-Gapped Switch Fabrics & Microburst Recovery](https://terminalledgerdesk.com/reports/multicast-a-b-market-data-feed-arbitration-switch-fault-isolation)
- Site research index: [Terminal Ledger Desk](https://terminalledgerdesk.com/llms-full.txt)
