Internet-Draft AI Training Fabric Benchmarking August 2026
Calabria, et al. Expires 13 February 2027 [Page]
Workgroup:
BMWG
Internet-Draft:
draft-calabria-bmwg-ai-fabric-training-bench-04
Published:
Intended Status:
Informational
Expires:
Authors:
F. Calabria
Cisco
C. Pignataro
Blue Fern Consulting
Q. Wu
Huawei
G. Fioccola
Huawei
S. Reddy
Apple

Benchmarking Methodology for AI Training Network Fabrics

Abstract

This document defines benchmarking terminology, methodologies, and Key Performance Indicators (KPIs) for evaluating Ethernet-based AI training network fabrics.

As large-scale distributed Artificial Intelligence / Machine Learning (AI/ML) training clusters grow to tens of thousands of accelerators (GPUs or generic accelerator processing units (XPUs)), the backend network fabric determines Job Completion Time (JCT), training throughput, and accelerator utilization.

This document establishes vendor-independent, reproducible test procedures for benchmarking fabric-level performance under realistic AI training workloads. The tests cover Remote Direct Memory Access (RDMA) over Converged Ethernet version 2 (RoCEv2) transport, the Ultra Ethernet Transport (UET) protocol defined by the Ultra Ethernet Consortium (UEC) Specification 1.0, congestion management (Priority Flow Control (PFC), Explicit Congestion Notification (ECN), Data Center Quantized Congestion Notification (DCQCN), Credit-Based Flow Control (CBFC)), load balancing strategies (Equal-Cost Multi-Path (ECMP), Dynamic Load Balancing (DLB), packet spraying), collective communication patterns (AllReduce, AllToAll, AllGather), and scale/soak testing.

The methodology enables direct, reproducible comparison across switch ASICs, NIC transport stacks (RoCEv2 and UET), and fabric architectures (2-tier Clos, 3-tier Clos, and rail-optimized).

About This Document

This note is to be removed before publishing as an RFC.

The latest revision of this draft can be found at https://fcalabri.github.io/bmwg-ai-fabric-training-bench/draft-calabria-bmwg-ai-fabric-training-bench.html. Status information for this document may be found at https://datatracker.ietf.org/doc/draft-calabria-bmwg-ai-fabric-training-bench/.

Discussion of this document takes place on the BMWG Working Group mailing list (mailto:bmwg@ietf.org), which is archived at https://mailarchive.ietf.org/arch/browse/bmwg/. Subscribe at https://www.ietf.org/mailman/listinfo/bmwg/.

Source for this draft and an issue tracker can be found at https://github.com/fcalabri/bmwg-ai-fabric-training-bench.

Status of This Memo

This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.

Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet-Drafts is at https://datatracker.ietf.org/drafts/current/.

Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."

This Internet-Draft will expire on 13 February 2027.

▲

Table of Contents

1. Introduction

Distributed AI/ML training workloads impose traffic requirements that standard data center fabrics were not designed to meet. Traditional data center traffic varies in flow size and protocol mix. AI training generates synchronized, bandwidth-intensive east-west traffic dominated by collective communication operations: AllReduce, AllToAll, and AllGather. These workloads require RDMA transport with negligible application-visible loss, bounded tail latency, uniform load distribution across all fabric paths, and the ability to absorb coordinated micro-bursts from thousands of accelerators simultaneously. RoCEv2 deployments typically meet the loss requirement by operating the fabric lossless (PFC/ECN); UET is designed to tolerate wire-level loss and recover via retransmission and packet trimming (see the Zero Packet Loss definition in [TERMINOLOGY]).

Existing BMWG methodologies do not address AI training fabrics. [RFC2544] defines benchmarking for general network interconnect devices but does not account for RDMA transport semantics, collective communication patterns, or the congestion behavior specific to GPU-to-GPU traffic. [RFC8238] and [RFC8239] establish data center benchmarking terminology and methodology but predate large-scale RoCEv2 deployment and do not address Priority Flow Control (PFC) interactions, DCQCN congestion control convergence [DCQCN-PAPER], or the impact of load balancing strategies on Job Completion Time (JCT). Industry experience deploying RoCEv2 at scale [META-ROCE] shows the need for a standardized benchmarking methodology.

The Ethernet Virtual Private Network (EVPN) benchmarking methodology [EVPN-BENCH] provides a structural template for service-oriented benchmarking but is scoped to L2VPN services rather than RDMA fabrics.

This document defines a benchmarking methodology for AI training network fabrics.

1.1. Requirements Language

The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "NOT RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in BCP 14 [RFC2119] [RFC8174] when, and only when, they appear in all capitals, as shown here.

1.2. Scope and Applicability

This document applies to Ethernet-based AI training backend network fabrics employing RoCEv2 and/or UEC Ultra Ethernet Transport (UET) protocols. The scope includes leaf-spine (2-tier Clos) and leaf-spine-superspine (3-tier Clos) topologies.

InfiniBand fabrics are explicitly out of scope, though many KPIs defined herein may be adapted for IB benchmarking by future documents. The DUT is the network fabric itself (the collection of switches and interconnecting links), not individual accelerators or host NICs; host-side configuration is documented in the test report as it materially affects results.

The DUT boundary for all measurements in this document is the NIC-to-NIC Ethernet fabric segment. Intra-node communication (proprietary accelerator interconnects, e.g., NVLink, Infinity Fabric/xGMI, or PCIe) and individual GPU/accelerator performance are explicitly out of scope. Collective operation measurements (AllReduce, AllGather, AllToAll) are measured at the Ethernet fabric boundary; intra-node accelerator-interconnect contributions are reported separately when characterizing wide Expert Parallelism (wide-EP) or multi-node configurations. The reporting requirements that make this separation auditable, and the normalization basis that keeps it intact, are specified in Section 3.2.

The methodology is designed for controlled laboratory environments per the BMWG charter; it is NOT intended for production network measurement.

1.3. Relationship to Existing and Companion Work

Table 1: Relationship to Existing and Companion Work
Document Relationship
[RFC1242] Base terminology for network benchmarking; terms reused herein
[RFC2544] Base methodology; throughput/latency/loss tests adapted for RDMA
[RFC8238] Data center terminology; buffer, congestion, and microburst terms extended
[RFC8239] Data center methodology; line-rate and buffer tests adapted for RoCEv2
[RFC9004] Back-to-back frame updates; burst absorption methodology referenced
[LLM-BENCH] Complementary document benchmarking the inference serving stack. Treats the network as opaque SUT; this document benchmarks the fabric itself. The two documents MAY be used together but MUST NOT be combined in a single benchmarking report without explicit section demarcation.
[TERMINOLOGY] Companion document; normative source for the terminology used throughout this document
[INFERENCE-BENCH] Companion fabric-level benchmarking methodology addressing AI inference serving workloads
[UEC-1.0] UET protocol specification; transport services, congestion control, and link-layer enhancements benchmarked in Section 6

2. Terminology

Terminology used in this document is defined in [TERMINOLOGY]. Readers should consult that document before applying the methodology defined here. Where a term overlaps with [RFC1242] or [RFC8238], the terminology document provides AI fabric context extensions; the foundational definitions in those RFCs remain authoritative for general network benchmarking.

All terminology used in this document – including the AI fabric, RoCEv2, UET, RDMA transport, congestion control (PFC, DCQCN, ECN, CBFC), load balancing (ECMP, Packet Spray, DLB/Flowlet), collective communication, and KPI vocabulary (JCT, JCT Ratio, BusBW, MMR, etc.) – is defined normatively in [TERMINOLOGY] and is not redefined here. The following table lists the single bench-specific extension introduced by this document:

Table 2: Bench-Specific Terminology Extensions
Term Definition
PFC Pause Event A single PFC PAUSE frame with a non-zero quanta value transmitted on a priority class. Zero-quanta (X-ON / resume) frames are excluded from the count, since implementations that signal resume explicitly would otherwise report roughly double the event rate of implementations that let the pause timer expire; X-ON frames are instead used to bound the paused interval for the cumulative-duration metric. Used in this document as the unit of count for PFC event-rate metrics (events/sec, cumulative duration) reported by the methodology in Section 7.

In addition to the BusBW reporting requirements specified in [TERMINOLOGY], the runtime algorithm selected by the collective library MUST be verified via library tracing and documented as part of the test conditions for any AllReduce, AllGather, or AllToAll benchmark in this document.

The scope of the DUT for the tests defined in this document is the set of leaf switches, spine switches, superspine switches (if applicable), and interconnecting links forming the AI training fabric, consistent with the Fabric DUT Boundary defined in [TERMINOLOGY].

2.1. Acronyms

Acronyms used in this document are expanded in the Acronyms appendix of [TERMINOLOGY]. Acronyms unique to the methodology defined herein are expanded on first use in the body of this document.

3. Test Topology and Architecture

3.1. Reference Fabric Topologies

Three reference topologies are defined. Every test report identifies which topology was used. Results obtained under different topologies are compared only under the conditions specified in Section 3.2, which defines the normalization basis used by this methodology and the parameters that MUST match before two results are compared.

3.1.1. Topology A: 2-Tier Clos (Leaf-Spine) 

+--------+ +--------+ +--------+ +--------+
| Spine1 | | Spine2 | | Spine3 | | SpineN |
+---++---+ +---++---+ +---++---+ +---++---+
    ||          ||          ||          ||
    ||    Full Mesh Interconnect        ||
    ||    (ECMP / DLB / Spray)         ||
    ||          ||          ||          ||
+---++---+ +---++---+ +---++---+ +---++---+
| Leaf 1 | | Leaf 2 | | Leaf 3 | | Leaf N |
+---++---+ +---++---+ +---++---+ +---++---+
    ||          ||          ||          ||
[GPU/XPU]  [GPU/XPU]  [GPU/XPU]  [GPU/XPU]
Hosts w/   Hosts w/   Hosts w/   Hosts w/
RoCEv2 NIC             RoCEv2 NIC
Figure 1: Topology A: 2-Tier Clos (Leaf-Spine)

The DUT boundary encompasses all leaf and spine switches and their interconnecting links. Traffic generators or actual GPU hosts connect at the leaf layer.

3.1.2. Topology B: 3-Tier Clos (Leaf-Spine-Superspine)

For clusters exceeding thousands of accelerators, a superspine layer is added. Each pod consists of a leaf-spine fabric; pods interconnect via superspine switches. This topology scales to 32,000+ accelerators at 800GbE with current-generation ASICs. The DUT boundary encompasses all three tiers.

3.1.3. Topology C: Rail-Optimized 

                       SPINE LAYER
+--------+ +--------+ +--------+ +--------+
| Spine1 | | Spine2 | | Spine3 | | SpineN |
+--+--+--+ +--+--+--+ +--+--+--+ +--+--+--+
 |     Full Mesh Interconnect (ECMP/Spray)  |
+--+--+--+ +--+--+--+ +--+--+--+ +--+--+--+
| Rail-0 | | Rail-1 | | Rail-2 | | Rail-7 |  RAIL (LEAF) LAYER
|  Leaf  | |  Leaf  | |  Leaf  | |  Leaf  |  one switch per NIC
+--+--+--+ +--+--+--+ +--+--+--+ +--+--+--+
  |   |       |   |       |   |      |   |
NIC-0 NIC-0 NIC-1 NIC-1 NIC-2 NIC-2 NIC-7 NIC-7
  |   |       |   |       |   |      |   |
+--------+ +--------+ +--------+ +--------+
| Host A | | Host B | | Host C | | Host D |  GPU HOSTS
| GPU[0] | | GPU[0] | | GPU[0] | | GPU[0] |  (each host has
| GPU[1] | | GPU[1] | | GPU[1] | | GPU[1] |   8 NICs, one
|  ...   | |  ...   | |  ...   | |  ...   |   per rail)
| GPU[7] | | GPU[7] | | GPU[7] | | GPU[7] |
+--------+ +--------+ +--------+ +--------+
Figure 2: Topology C: Rail-Optimized (schematic; only 4 of N rails and hosts shown)

In rail-optimized topologies, each NIC on a multi-NIC host connects to a dedicated leaf switch ("rail"); this co-optimizes network locality with the collective communications library (CCL) in use (e.g., NCCL, RCCL, oneCCL). The diagram is schematic: it depicts a representative subset of rails and hosts, not the full cross-host fan-out; in the complete topology, the Rail-N leaf switch connects to GPU[N] of every host. The DUT boundary and rail mapping are fully documented in the test report.

3.2. Result Comparability and Normalization

Section 3.1 requires that results obtained under different topologies be compared only after normalization. This section defines what that normalization is, which parameters MUST be equal before two results are compared, and which quantities MUST NOT be used as normalization factors.

3.2.1. Normalization Basis

Normalization in this document is achieved by reporting KPIs that are dimensionless or referenced to a rate observable at the Fabric DUT Boundary defined in [TERMINOLOGY]. It is not achieved by normalizing fabric hardware between the systems being compared.

Table 3: Normalization Basis for Reported KPIs
KPI Defined in Normalizing denominator
JCT Ratio Section 10.1 Roofline_seq, whose communication term is (8 × S × algo_factor) / B_acc
BusBW Section 9 Per-accelerator; made algorithm-invariant by the fixed algo_factor of [TERMINOLOGY]
BusBW efficiency Section 9.1, Section 9.3 Per-accelerator NIC line rate
Throughput efficiency Section 5.1 Theoretical rate of the Ethernet links under test
Aggregate Throughput Section 4 Fabric bisection bandwidth, computed from Ethernet link rates
MMR, JFI Section 8 Dimensionless by construction
JCT Interference Factor Section 10.3 Baseline JCT measured on the same DUT

Each denominator in Table 3 is either a workload parameter fixed by the test operator (S, algo_factor, N) or an Ethernet line rate at the DUT boundary. No denominator contains a term for switch internal capacity, accelerator interconnect capacity, or host bus capacity.

Convergence and failover times (Section 11.2, Section 11.3), latency percentiles (Section 5.2), and queue occupancy are reported in absolute units and are not normalized; normalizing them would obscure the behavior they measure.

3.2.2. Aggregate Switching Capacity Is Not a Normalization Factor

Aggregate switching capacity (ASIC forwarding capacity, in Tbps) MUST NOT be used as a normalization factor for any KPI defined in this document, and MUST NOT be held constant as a precondition for comparing two fabrics, for three reasons:

  1. Aggregate switching capacity is a vendor-declared internal property, not a quantity observable at the Fabric DUT Boundary. Benchmarking in this document is performed on a black-box basis (Section 14).

  2. Whether a given switching capacity is sufficient for the offered collective pattern is what the tests in Section 5 through Section 12 measure. Dividing a measured result by the capacity that produced it removes the effect under test: a fabric with half the switching capacity that completes the same job in 1.6 times the time would report a better result per Tbps while performing worse at the job.

  3. A capacity-normalized value is a cost or efficiency figure of merit. Per the BMWG charter, and consistent with Section 13 and Appendix B, this document defines what is measured and how it is reported, and does not define figures of merit or acceptance criteria.

Aggregate switching capacity, per-port speed, switch radix, and buffer architecture are reported as DUT characteristics (Section 3.3, Appendix C) so that a reader can attribute an observed difference to them. They are inputs to the interpretation of a result, not divisors applied to it.

The same reasoning applies to intra-node switching and interconnect capacity (accelerator interconnects, PCIe, CXL). Were aggregate switching capacity used as a normalization factor, intra-node switching and interconnect capacity would have to be included for that normalization to be complete, which the DUT boundary of Section 1.2 does not permit. Because no denominator in Table 3 contains such a term, the normalization defined here does not reach inside the node, and this section and Section 1.2 are consistent by construction.

3.2.3. Comparability Set

Results obtained on two different fabrics are compared directly for the KPIs in Table 3 when every parameter in Table 4 is equal and reported for both results.

Table 4: Comparability Set
Parameter Why it must match
Participating accelerator count N Collective cost depends on N through algo_factor and through fabric path length
Accelerators per node, and their placement and rail mapping Determines the fabric-visible fraction of collective traffic (Section 3.2.5)
B_acc, the per-accelerator Ethernet line rate at the DUT boundary Denominator of Roofline_seq and of BusBW efficiency
Leaf oversubscription ratio Determines whether the fabric can satisfy the offered pattern
Collective type, message size S, and algo_factor Definition of the offered workload
Transport (RoCEv2 or UET), and for UET the transport service Loss and retransmission semantics differ by transport
Load balancing strategy Comparisons are made strategy by strategy

When any parameter in Table 4 differs, the results are reported side by side with the difference stated, and are not combined into a single comparative figure.

Topology class is a reported test condition, not a parameter that normalization removes. Two fabrics of different topology class that match on Table 4 are compared for the KPIs in Table 3, and the report states, for each result: the topology class; the switch tier count; the typical and worst-case hop count for the measured traffic pattern; the number of equal-cost paths available between a source-destination pair; and the bisection bandwidth. Such a comparison characterizes the topologies under the offered workload; it does not attribute the observed difference to any single fabric component.

Topologies outside the reference set of Section 3.1, including mesh, torus, and direct-connect fabrics, are outside the scope stated in Section 1.2. Results obtained on such a fabric are reported as a deviation per Section 13, and the hop-count and equal-cost-path descriptors above are supplied only where they are well defined for that topology.

3.2.4. Comparing Fabrics of the Same Topology Class

Different scale. Comparison is performed at equal N, and scale is characterized rather than normalized away. Section 10.1 requires JCT Ratio to be reported for each N in its parameter table and plotted against N; the shape of that curve is the scaling result, and no single scalar replaces it. Two fabrics whose maximum supported scale differs are compared over the range of N that both support; the report gives the JCT Ratio at the largest common N together with the maximum N each fabric sustains per Section 11.1.

Different switch capacity, radix, or buffer size. Comparison is performed without adjustment, provided Table 4 matches. The difference in the measured KPIs is the result of the comparison. The differing DUT characteristics are reported per Section 3.3 and Appendix C to permit attribution, and are not used to scale the measured values.

3.2.5. Fabric-Visible Data Volume

Collective placement determines how much of a collective's traffic crosses the DUT boundary. A hierarchical AllReduce that reduces within a node before reducing across the fabric presents less data to the fabric than a flat AllReduce across the same N accelerators, even though the application-level message size S is identical. Two results obtained under different placement therefore represent different offered fabric workloads at equal S.

For every collective result reported under Section 9, Section 6.6, or Section 10, the report states:

  • S, the application-level data size per participant;

  • S_fabric, the data volume per participant that crosses the Fabric DUT Boundary, together with the method used to obtain it. S_fabric is counted in application payload bytes, each byte counted once, using the byte-counting rule of the Fabric_Goodput definition in [TERMINOLOGY]. Retransmitted and duplicate bytes MUST NOT be included in S_fabric; they are reported separately as the RDMA Retransmission Rate and out-of-order delivery rate of Section 4. Measurement from NIC Ethernet port counters is preferred; derivation from the collective algorithm and the placement is acceptable when the derivation is stated. Under this byte-counting rule the two methods return the same value; when both are available and they disagree, the report gives both values and the difference;

  • the accelerator placement and, in rail-optimized topologies, the rail mapping;

  • the Intra-Node Transfer Overhead component defined in [TERMINOLOGY], reported separately. This component is never added to, subtracted from, or folded into a fabric KPI.

Comparisons use the same S and the same placement on both sides. When placement cannot be matched, for example when comparing a rail-optimized fabric using rail-aware placement against a Clos fabric that has no rail structure, the report gives S_fabric for each result so that the difference in offered fabric work is visible, and the results are not presented as an equal-workload comparison.

3.3. Device Under Test (DUT) Identification

Table 5: DUT Identification Parameters
Parameter Description Example
Switch Vendor/Model Vendor name, product family, model number Vendor Family Model
Switch ASIC Silicon vendor, ASIC family, revision Silicon Vendor ASIC Family Rev
NOS Version Network operating system name and version NOS Name Version
Port Speed Per-port line rate 400GbE, 800GbE
Buffer Architecture Shared/dedicated, total buffer per ASIC/port 32MB shared + 16MB VOQ per port
Optics/Cables Transceiver type, cable type and length Octal Small Form-factor Pluggable (OSFP) 400G-DR4, Direct Attach Copper (DAC) 3m cable
NIC Vendor/Model RDMA NIC vendor, model, firmware NIC Vendor Model Speed
NIC Firmware NIC firmware version Firmware Version
Host Config OS, CCL lib version, driver, BIOS settings OS Version, CCL Version, OFED Version

3.4. Traffic Generator Requirements

3.4.1. Mandatory Functional Capabilities

The traffic generator supports: RoCEv2 transport emulation (QP establishment, RDMA Write/Read, ECN processing, DCQCN rate control); configurable QP scaling (1-256 QPs per source-destination pair); programmable collective communication patterns (AllReduce, AllToAll, AllGather with configurable message sizes); and nanosecond-precision timestamping.

3.4.2. Minimum Measurement Accuracy Requirements

Table 6: Minimum Measurement Accuracy Requirements
Parameter Minimum Requirement
Timestamp accuracy ≤ 100 nanoseconds
Frame rate accuracy +/- 0.1% of specified rate
QP scaling range 1 to 256 QPs per src-dst pair
Message size range 64 B to 2 GB (single-message ceiling; see note)
Flow counter resolution Per-flow byte and packet counts
Loss measurement Exact per-packet loss counting
Burst generation Burst lengths at line rate sufficient to exceed DUT buffering; configurable beyond 1000 frames

NOTE: A single RDMA message cannot exceed the 32-bit RETH DMA Length field, an absolute ceiling of 4 GB, and implementations commonly advertise a practical max_msg_sz of 1-2 GB. Transfers larger than the single-message ceiling are composed from multiple RDMA messages, and the message count and per-message size are reported alongside the aggregate transfer size.

3.4.3. Acceptable Implementations

The platform used is identified in all test reports.

(a) Hardware Traffic Generator – dedicated hardware capable of line-rate RDMA emulation meeting the Measurement Accuracy Requirements specified in this document. Suitable for point-to-point RDMA tests (Section 5 and Section 6). For collective tests (Section 9), the following limitations are documented: whether synchronization barriers are reproduced, whether flow patterns are schedule-driven or gradient-driven, and whether straggler behavior is modeled.

(b) Accelerator Cluster – cluster running an actual collective communication library with RDMA tooling. Preferred for the collective benchmarks in Section 9. Host configuration (accelerator model, collective library name and version, PCIe topology, BIOS power management settings) is documented. Any non-fabric overhead in timing measurements is quantified and reported separately.

When a hardware generator is used for collective benchmarks, results should be cross-validated against an accelerator cluster at one or more overlapping (message_size, N) configurations.

Discrepancies exceeding 10% in BusBW or JCT Ratio are investigated and reported.

4. KPI Framework and Metrics Taxonomy

4.1. Primary KPIs

Table 7: Primary KPIs
KPI Unit Definition
Job Completion Time (JCT) seconds Wall-clock time for benchmark iteration (compute + communication)
JCT Ratio dimensionless Measured JCT / Roofline JCT
Bus Bandwidth (BusBW) Gbps/accelerator Effective per-accelerator throughput during collective. See the BusBW definition in [TERMINOLOGY]
Aggregate Throughput Tbps Total fabric goodput during collective phase
Packet Drop Rate ppm Frames lost end-to-end not retransmitted
Tail Latency (P99/P99.9) us 99th/99.9th percentile one-way fabric latency

4.2. Secondary KPIs

Table 8: Secondary KPIs
KPI Unit Definition
ECN Marking Ratio % Percentage of packets marked CE over measurement interval
PFC Pause Count events/sec Rate of PFC PAUSE frames per priority per port
PFC Pause Duration us Cumulative time a port is in PFC-paused state per interval
RDMA Retransmission Rate retx/sec NIC-level retransmissions due to timeouts or NAKs
ECMP Imbalance (MMR) dimensionless Max-Mean Ratio of flow counts across parallel uplinks
Jain's Fairness Index (JFI) 1/N-1.0 Fairness of traffic distribution; 1.0 = perfect, 1/N = worst (N = number of parallel links)
Queue Depth (P95/Max) bytes or cells 95th percentile and maximum egress queue occupancy per port
Congestion Control Convergence us Time from congestion onset to DCQCN rate stabilization
Out-of-Order Packet Rate pkt/sec Packets delivered out of sequence (relevant for packet spray)

4.3. Fabric Health Indicators

Table 9: Fabric Health Indicators
Indicator Unit Definition
Switch CPU Utilization % Average and peak CPU usage on DUT control plane during test
Switch Memory Utilization % Average and peak memory usage, including FIB/MAC table occupancy
Forwarding Information Base (FIB) / Route Convergence Time ms Time to converge routing after topology change
Link Flap Count events Spurious link state changes during test period
CRC/FCS Error Rate errors/sec Physical layer errors indicating cable or optics issues
Power Consumption Watts Per-switch and per-port power draw under test load

5. Test Category 1: RDMA Transport Benchmarks

These tests establish baseline fabric performance for RDMA traffic independent of collective communication patterns. They extend [RFC2544] and [RFC8239] methodology for RoCEv2 semantics.

5.1. Baseline Throughput

Objective: Determine the maximum sustainable RDMA Write throughput through the DUT fabric at each tested message size.

Procedure:

  • Configure N host pairs, each establishing Q Queue Pairs per pair

  • Initiate RDMA Write operations and measure aggregate goodput

  • Each test runs for at least 60 seconds at each rate

  • Binary search per [RFC2544] Section 26.1 is used

  • Message sizes: 64B, 256B, 1KB, 4KB, 64KB, 256KB, 1MB, 4MB

  • QP counts: 1, 4, 16, 32 per src-dst pair

  • Test both unidirectional and bidirectional traffic

Reporting: Report aggregate throughput (Tbps), per-port utilization (%), and throughput efficiency (measured/theoretical). Present as table indexed by message size × QP count, and as graph (message size on X-axis).

5.2. Latency Characterization

Objective: Determine one-way and round-trip RDMA latency distribution at the throughput rate from Section 5.1.

Procedure:

  • Inject tagged frames at 60s into a 120s stream (per [RFC2544] Section 26.2)

  • Nanosecond-precision timestamping

  • Reported statistics: min, mean, P50, P95, P99, P99.9, max

  • Each run MUST capture at least 10,000 latency samples to support a statistically meaningful P99.9

  • Repeat at least 20 times; pool all samples across runs and compute percentiles from the pooled distribution, and report the per-run variability (e.g., min/max across runs) alongside the pooled statistics

  • Test under both zero-load (single QP) and loaded (full fabric utilization) conditions

Reporting: Tabulate latency statistics per message size. Provide histogram and CDF plot. Report latency increase factor (loaded/unloaded).

5.3. Back-to-Back Burst Absorption

Objective: Characterize the DUT fabric's ability to absorb back-to-back RDMA bursts without loss. This test extends [RFC9004] methodology for RoCEv2.

Procedure:

  • Transmit bursts at line rate with minimum inter-frame gap

  • Increase burst length until first frame loss is detected

  • Test incast ratios: 2:1, 4:1, 8:1, 16:1, 32:1

  • Repeat at least 50 times per burst length

Reporting: Report burst absorption capacity (frames and bytes) for each message size and incast ratio. Plot burst capacity vs. incast ratio.

6. Test Category 2: UEC Transport Protocol Benchmarks

The Ultra Ethernet Consortium (UEC) Specification 1.0 [UEC-1.0] defines UET, an RDMA transport positioned as an alternative to RoCEv2 for AI/HPC workloads. All UET tests use the libfabric API [LIBFABRIC] and run on UEC 1.0-compliant NICs.

The UEC compliance profile (AI Base, AI Full, or HPC) used during testing is documented in the test report.

6.1. UET Throughput by Transport Service

Objective: Determine maximum sustainable throughput under each UET transport service (ROD, RUD, RUDI, UUD) and compare to RoCEv2 Reliable Connected (RC) / Unreliable Connected (UC) on the same DUT fabric.

Procedure: Use UEC 1.0-compliant NICs; establish PDCs; use libfabric fi_write. Apply binary search ([RFC2544] Section 26.1). Vary PDC counts: 1, 4, 16, 32. A parallel RoCEv2 test series is executed for comparison. Both unidirectional and bidirectional configurations are tested.

Reporting template:

Table 10: UET Throughput by Transport Service
Metric ROD RUD RUDI UUD RoCEv2 RC RoCEv2 UC
Throughput @ 1MB (Gbps) (meas) (meas) (meas) (meas) (meas) (meas)
Throughput @ 4MB (Gbps) (meas) (meas) (meas) (meas) (meas) (meas)
Efficiency (% line rate) (meas) (meas) (meas) (meas) (meas) (meas)
Connection/PDC Initiation Latency (us) (see note) (meas) (meas) (meas) (meas) (meas) (meas)
Max Sustained PDC/QP Count (meas) (meas) (meas) (meas) (meas) (meas)

NOTE: For RoCEv2 RC/UC, this measures QP connection setup time (handshake round-trip). For UET (ROD/RUD/RUDI/UUD), PDCs have no separate setup handshake; this measures the first-packet latency reflecting in-band PDC initiation instead.

6.2. UET Latency Characterization

Objective: Measure latency distribution for UET transport services; quantify differential vs. RoCEv2, with particular attention to connectionless PDC establishment overhead.

Procedure: Measure latency for: (a) steady-state PDC transfers; (b) first-packet latency (PDC + first data packet, measuring "data before handshake"); (c) zero-load baseline. Test ROD and RUD separately to isolate reordering-related latency.

Reporting: Tabulate latency statistics per (transport_service, message_size, load_condition) tuple. Plot latency CDF for UET ROD, UET RUD, and RoCEv2 RC side-by-side.

6.3. Packet Spray Efficacy Under UET RUD

Objective: Quantify the load balancing improvement achieved by UET's native per-packet spray with RUD, which eliminates the receiver reorder buffer constraint.

Procedure: Test five configurations:

  • UET RUD + packet spray

  • UET ROD + packet spray

  • RoCEv2 RC + packet spray

  • RoCEv2 RC + standard ECMP (baseline)

  • UET RUD + DLB/Flowlet

Measure MMR, JFI, out-of-order delivery rate, retransmission rate, and effective goodput. Vary ECMP paths: 4, 8, 16, 32.

Reporting template:

Table 11: Packet Spray Efficacy Under UET RUD
Load Balancing Config MMR JFI OOO Rate Retx Rate Effective Goodput (%)
UET RUD + Packet Spray (meas) (meas) (meas) (meas) (meas)
UET ROD + Packet Spray (meas) (meas) (meas) (meas) (meas)
RoCEv2 RC + Packet Spray (meas) (meas) (meas) (meas) (meas)
RoCEv2 RC + ECMP (baseline) (meas) (meas) (meas) (meas) (meas)
UET RUD + DLB/Flowlet (meas) (meas) (meas) (meas) (meas)

  • This test evaluates whether UET RUD achieves zero host-visible reordering despite per-packet spray, since the transport layer is designed to tolerate unordered delivery.

6.4. UET Congestion Control Benchmarks

Objective: Evaluate UET's dual-sided (sender + receiver) congestion control under N:1 incast conditions vs. RoCEv2 DCQCN.

Procedure: Measure: (a) incast throughput at N = {2, 4, 8, 16, 32, 64}; (b) convergence time after doubling active senders (until all flows within 10% of fair share); (c) PFC avoidance with PFC disabled on the DUT; (d) receiver credit utilization.

Reporting: Tabulate incast throughput, convergence time, peak queue depth, PFC event count, and packet drop rate for UET vs. DCQCN per incast ratio. Key metric: report whether UET achieves zero application-visible loss without PFC.

6.6. UET Collective Communication Performance

Objective: Measure collective communication (AllReduce, AllToAll, AllGather) performance over UET and compare directly to RoCEv2, isolating the transport protocol contribution to collective efficiency.

Procedure: Execute the collective benchmark suite from Section 9 over UET RUD transport using a UEC-compliant collective library. The same accelerator count (N), message sizes, and fabric topology are used for both UET and RoCEv2 runs to ensure a valid comparison. Run UET RUD + packet spray as the primary configuration and UET ROD + ECMP as the secondary baseline.

For AllReduce, the UET TSS group-key encryption state (active or inactive) on the DUT NIC is documented as a required result field in the test report.

When UET TSS group-key encryption is active during testing, report the observed BusBW computed from measured bytes transferred per the algo_factor formula defined in [TERMINOLOGY] (fixed per collective type); group-key encryption affects per-packet security processing overhead, not the transfer volume itself.

The runtime algorithm in use is reported per message-size bucket. See [TERMINOLOGY] for the BusBW definition and algo_factor values.

Reporting: Report the percentage improvement in BusBW and JCT attributable to UET native packet spray and congestion control.

Reporting template:

Table 12: UET Collective Communication Performance
Collective Msg Size N Accels UET RUD BusBW UET ROD BusBW RoCEv2 RC BusBW Delta UET/RoCEv2
AllReduce 1GB 128 (meas) (meas) (meas) (meas)
AllReduce 1GB 512 (meas) (meas) (meas) (meas)
AllToAll 1GB 128 (meas) (meas) (meas) (meas)
AllGather 1GB 128 (meas) (meas) (meas) (meas)

6.7. UET PDC Scalability and Connection Setup Rate

Objective: Measure PDC establishment rate and maximum concurrent PDC count vs. RoCEv2 QP-based connections.

Procedure: (a) PDC establishment rate: initiate PDC creation to M = {100, 1000, 10000, 100000} remote endpoints. (b) Data-before-handshake: measure first-byte latency for UET vs. RoCEv2 RDMA Write. (c) Maximum concurrent PDC count: scale until per-PDC throughput drops below 90% of single-PDC rate. The UEC specification [UEC-1.0] targets millions of endpoints.

NOTE: PDC and QP scaling limits are a host NIC capability, outside the DUT boundary per [TERMINOLOGY] and Section 1.2. They are reported as context because a NIC-side PDC/QP ceiling presents as fabric underperformance in the throughput and latency measurements above.

7. Test Category 3: Congestion Management

AI training workloads generate repetitive micro-congestion during the back-propagation gradient synchronization phase.

7.1. ECN Marking Accuracy and Threshold

Objective: Verify that the DUT marks packets with ECN CE at the configured threshold with correct granularity.

Procedure: Configure threshold T on DUT egress queue. Verify: (a) no packets marked below T; (b) 100% marked above maximum threshold; (c) appropriate Weighted Random Early Detection (WRED) / Random Early Detection (RED) probability ramp between thresholds. Test thresholds: low (~100KB), medium (~1MB), high (~5MB).

Reporting: Plot ECN marking probability vs. instantaneous queue depth. Report measured threshold accuracy (deviation from configured).

7.2. PFC Behavior Under Incast

Objective: Characterize DUT's PFC generation behavior under N:1 incast conditions.

Procedure: Generate N:1 incast at 100% line rate, N = {2, 4, 8, 16, 32, 64}. Measure PFC PAUSE frame count/sec per hop, PFC PAUSE duration per port, PFC storm onset, and end-to-end throughput. The test characterizes headroom sizing and PFC watchdog effectiveness.

7.3. DCQCN Convergence Time

Objective: Measure time for DCQCN to converge to fair-share rate after congestion onset.

Procedure: Establish M flows through a common bottleneck. At T0, inject additional M flows (creating 2:1 oversubscription). Measure time until all 2M flows achieve rates within 10% of fair share. Repeat for M = {4, 16, 64, 256}. Vary DCQCN parameters and report sensitivity.

7.4. PFC Storm and Deadlock Resilience

Objective: Verify the DUT does not enter PFC deadlock or sustained PFC storm under adversarial traffic.

Procedure: Generate cyclic traffic patterns known to cause PFC deadlocks. Run for 300 seconds. The test characterizes whether the DUT demonstrates resilience via PFC watchdog or architectural immunity (e.g., VOQ-based scheduling); the mechanism observed is reported.

8. Test Category 4: Load Balancing Efficacy

Load balancing across parallel fabric paths is critical for AI training fabrics because the traffic consists of a small number of high-bandwidth, long-lived elephant flows.

8.1. ECMP Entropy and Polarization

Objective: Quantify traffic polarization under standard ECMP hashing for AI training flow patterns.

Procedure: Configure standard 5-tuple ECMP. Generate traffic with Q = {1, 4, 8, 16, 32} QPs per src-dst pair. Measure per-link utilization, MMR, and JFI. Test with and without ECMP hashing that includes the BTH destination QP field as a hash input (in addition to the standard 5-tuple). Repeat for fabric sizes of 8, 16, 32, and 64 leaf switches.

8.2. Dynamic Load Balancing (Flowlet)

Objective: Evaluate DUT's flowlet-based DLB performance and compare to baseline ECMP.

Procedure: Configure vendor-specific DLB (document algorithm type). Generate traffic with Q=4 QPs. Measure MMR, JFI, per-link utilization, out-of-order rate. Vary flowlet gap timer and report sensitivity.

8.3. Packet Spraying

Objective: Evaluate DUT's per-packet spraying performance and quantify the utilization vs. reordering tradeoff.

Procedure: Configure per-packet load balancing. Measure MMR (expected ~1.0), JFI (expected ~1.0), out-of-order rate, and RDMA retransmission impact. If the DUT provides an in-fabric reorder buffer, document per Appendix C.

8.4. Jain's Fairness Index Measurement

Objective: Single-number summary of load balancing quality comparable across all strategies.

Formula:

JFI = (Sum LinkTx_i)^2 / (N × Sum LinkTx_i^2)
Figure 3: Jain's Fairness Index Formula

where LinkTx_i = transmitted traffic on fabric link i, N = total parallel links. Range: 1/N (worst) to 1.0 (perfect).

Reporting: Report JFI for each load balancing strategy. Provide bar chart comparing ECMP, DLB, and packet spray.

9. Test Category 5: Collective Communication Benchmarks

These tests evaluate the fabric's performance under realistic collective communication patterns. Unlike synthetic RDMA tests in Section 5 and Section 6, these exercise the full stack including the collective communications library (CCL) in use (e.g., NCCL, RCCL, oneCCL).

Because collective placement determines how much of a collective's traffic crosses the DUT boundary, every result in this section is reported together with the fabric-visible data volume and placement information required by Section 3.2.5.

9.1. AllReduce Benchmark

Objective: Measure fabric performance during AllReduce operations, the dominant collective for gradient synchronization in data-parallel training.

Procedure: Using N accelerators connected through the DUT fabric, execute AllReduce (sum) operations using a collective communications library benchmark suite (e.g., nccl-tests, rccl-tests, or equivalent).

Test parameters:

  • Message sizes: 1 MB, 8 MB, 64 MB, 256 MB, 1 GB, 4 GB

  • Accelerator counts (N): 8, 16, 32, 64, 128, 256, 512, 1024

  • Minimum iterations per (message_size, N) pair: 100

  • Load balancing strategies: ECMP, DLB, packet spray

For each (message_size, N) pair, record average, P50, P95, and P99 BusBW, ECN marking ratio, PFC pause count, and per-link utilization. BusBW is computed per the BusBW definition in [TERMINOLOGY]; algo_factor is fixed per collective type and does not vary with the algorithm the library selects at runtime. The runtime algorithm selected by the library for each message-size bucket is verified via library tracing and documented as part of the test conditions.

Reporting: Tabulate BusBW for each (message_size, N, LB_strategy, Algorithm (verified)) combination. The "Algorithm (verified)" column is required; results without it are incomplete. Plot BusBW vs. N for each message size. Report BusBW efficiency = BusBW / NIC_line_rate.

9.2. AllToAll Benchmark

Objective: Measure fabric performance during AllToAll operations, the dominant collective for Mixture-of-Experts (MoE) expert parallelism dispatch.

Procedure: Using the same message sizes, accelerator counts, iteration count, and load balancing strategies as Section 9.1, execute AllToAll operations via the collective communication library.

AllToAll generates the worst-case fabric stress pattern: every accelerator simultaneously sends a unique payload to every other accelerator in the group, which creates maximum entropy and stresses every fabric link with many-to-many traffic overlap. This makes AllToAll JCT the most sensitive single indicator of fabric congestion management quality.

BusBW is computed per the BusBW definition in [TERMINOLOGY]; algo_factor is fixed per collective type and does not depend on topology or library implementation. The runtime algorithm in use is verified via library tracing and documented as part of the test conditions.

Measurement: Report BusBW (average, P50, P95, P99), JCT per iteration, ECN marking ratio, PFC pause count, and per-link utilization for each (message_size, N, LB_strategy) combination.

Reporting: Same table format as Section 9.1, with the "Algorithm (verified)" column required. Additionally report JCT for each configuration; JCT degradation relative to the ECMP baseline is highlighted as the primary congestion sensitivity indicator.

9.3. AllGather Benchmark

Objective: Measure fabric performance during AllGather operations, the dominant collective for weight and activation distribution in tensor-parallel training.

Procedure: Using the same message sizes, accelerator counts, iteration count, and load balancing strategies as Section 9.1, execute AllGather operations via the collective communication library.

AllGather consists of a gather phase only – each accelerator contributes a shard and receives the full concatenated tensor. There is no reduce phase, which produces lower peak fabric load than AllReduce at equivalent message size and N. This makes AllGather a useful baseline for isolating the gather-path fabric contribution from the combined send-and-reduce cost.

BusBW is computed per the BusBW definition in [TERMINOLOGY]; algo_factor is fixed per collective type and does not depend on the library's algorithm selection. The runtime algorithm in use is verified via library tracing and documented as part of the test conditions.

Measurement: Report BusBW (average, P50, P95, P99), JCT per iteration, ECN marking ratio, PFC pause count, and per-link utilization for each (message_size, N, LB_strategy) combination.

Reporting: Same table format as Section 9.1, with the "Algorithm (verified)" column required. Report BusBW efficiency = BusBW / NIC_line_rate. Where results are compared to AllReduce under identical parameters, the BusBW ratio (AllGather / AllReduce) quantifies the difference in fabric load between the two traffic patterns; AllReduce's additional network traffic reflects its ReduceScatter phase, not the reduction arithmetic itself, which is performed by the accelerators, not the fabric.

9.4. Collective Communication Library Bus Bandwidth Summary

Reporting template:

Table 13: Collective Communication Bus Bandwidth Summary
Collective Msg Size N Accels ECMP BusBW (Gbps/accel) DLB BusBW (Gbps/accel) Spray BusBW (Gbps/accel)
AllReduce 1GB 128 (meas) (meas) (meas)
AllReduce 1GB 512 (meas) (meas) (meas)
AllToAll 1GB 128 (meas) (meas) (meas)
AllToAll 1GB 512 (meas) (meas) (meas)
AllGather 1GB 128 (meas) (meas) (meas)
AllGather 1GB 512 (meas) (meas) (meas)

10. Test Category 6: Job Completion Time (JCT) Benchmarks

JCT is the single most important user-facing KPI for AI training fabrics; it directly determines accelerator utilization and training cost.

10.1. Synthetic JCT Under Controlled Conditions

Objective: Measure JCT for a defined synthetic workload with a known computation-to-communication ratio to isolate fabric-induced overhead.

Procedure: Define a synthetic training iteration as a strictly sequential model:

  1. Computation phase of C milliseconds (simulated sleep or GPU compute kernel)

  2. Communication phase: AllReduce of S bytes across N accelerators

Table 14: Synthetic JCT Test Parameters
Parameter Values
Computation time C 10 ms, 50 ms, 100 ms, 500 ms
Message size S 256 MB, 1 GB, 4 GB
Accelerator count N 64, 128, 256, 512, 1024
Iterations 1000

Execute 1000 iterations and measure total wall-clock JCT.

Roofline_seq = Iterations × (C + (8 × S × algo_factor) / B_acc)
JCT Ratio    = Measured_JCT / Roofline_seq

  where:
    C            = compute time per iteration, in seconds
                   (convert the millisecond values in the
                   parameter table to seconds)
    S            = message size per iteration (bytes)
    algo_factor  = fixed normalization constant per collective
                   type; see the BusBW definition in the
                   companion terminology document
    B_acc        = aggregate per-accelerator NIC line rate
                   (bits/second); sum across all NICs serving
                   the accelerator (e.g., in rail-optimised
                   topologies, the sum of all rail NIC speeds)
    Iterations   = number of synthetic iterations executed

  The factor of 8 converts S from bytes to bits to match the
  units of B_acc.
Figure 4: JCT Ratio Calculation

This model assumes strictly sequential compute and communication phases and represents a conservative upper bound on communication overhead. Many frameworks overlap these phases via gradient bucketing or asynchronous collectives, which reduces the effective communication overhead visible in wall-clock JCT.

Implementations using overlapped execution additionally report:

Overlap_Fraction = 1 - (Measured_JCT - C_total) / Comm_time

  where:
    C_total   = Iterations × C
    Comm_time = Iterations × (8 × S × algo_factor) / B_acc
    S, algo_factor, B_acc as defined for Roofline_seq above.
Figure 5: Overlap Fraction Calculation

An Overlap_Fraction of 0 indicates fully sequential execution; 1.0 indicates communication is perfectly hidden behind compute.

When overlap is present, the residual fabric overhead is reported as:

Effective_Comm_Overhead = Measured_JCT - C_total

The Overlap_Fraction and communication-library overlap configuration (e.g., bucket size, number of async streams) are documented as part of the test configuration when this optional measurement is reported.

Reporting: Tabulate JCT Ratio for each (C, S, N, LB_strategy) combination. Plot JCT Ratio vs. N to characterize fabric scalability.

  • NOTE: JCT Ratio values of 1.05 and 1.15 are cited elsewhere in this document (Appendix B) as illustrative reference points, not as pass/fail thresholds. Per the BMWG charter, the definition of acceptance criteria or performance requirements is explicitly outside the scope of this Working Group; deployment-specific thresholds are outside the scope of this document.

10.2. MLPerf-Aligned JCT

Objective: Measure JCT using MLPerf Training benchmark workloads [MLPERF] to enable comparison with published industry results.

Procedure: Execute the current MLPerf Training closed-division workloads per MLPerf submission rules ([MLPERF]); the specific workload set changes across MLPerf versions, so the test report MUST identify the MLPerf Training version and workload names used. Simultaneously capture all fabric KPIs from Section 4. Report time-to-train and/or tokens-per-second.

10.3. Multi-Tenant JCT Interference

Objective: Quantify JCT impact when multiple training jobs share the same fabric.

Procedure: Configure two or more independent training jobs. Jobs are configured to overlap in spine-layer link usage. Measure baseline JCT (isolated) and contention JCT (simultaneous).

JCT Interference Factor = Contention_JCT / Baseline_JCT
Figure 6: JCT Interference Factor

Test with spine link overlap: 0%, 25%, 50%, 75%.

11. Test Category 7: Scale and Convergence

11.1. Fabric Scale Limits

Objective: Determine the maximum fabric scale at which the DUT maintains acceptable KPI performance.

Procedure: Progressively increase active accelerator endpoints from N=64 to maximum topology support while running AllReduce (Section 9.1, S=1GB). At each scale point record JCT Ratio, BusBW, ECN ratio, PFC count, CPU and memory utilization. Also measure BGP/routing convergence time after clearing all adjacencies (analogous to the convergence testing approach in [EVPN-BENCH]).

11.3. Zero-Impact Failover Measurement

Objective: Verify vendor claims of zero-impact or sub-microsecond failover.

Procedure: Execute Section 11.2 with nanosecond-precision measurement. A failure is considered "zero-impact" if the measured JCT for the failure iteration is within the P99 JCT of steady-state iterations.

12. Test Category 8: Soak and Stability

12.1. 24-Hour Sustained Load

Objective: Characterize DUT fabric stability under sustained AI training load over an extended period, following the soak-testing methodology pattern in [EVPN-BENCH].

Procedure: Configure DUT at maximum validated scale from Section 11.1. Generate bidirectional collective communication traffic (alternating AllReduce and AllToAll) at 80% of maximum validated throughput, per the offered load fraction required by the Soak Test definition in [TERMINOLOGY]. Run continuously for 24 hours. Sample all KPIs from Section 4 every 60 seconds.

The objective of the soak test is to monitor and document fabric behavior under extended load. The methodology does not establish pass/fail criteria for any reported metric. Any memory leaks, crashes, or other anomalies encountered during the test MUST be documented as an application log file or other dedicated file with their timestamps and durations.

Reporting: Time-series plots of JCT Ratio, BusBW, ECN ratio, PFC count, CPU, and memory over the 24-hour period. Report standard deviation of JCT Ratio (stability metric).

12.2. Resource Leak Detection

Objective: Detect memory leaks, handle exhaustion, or gradual performance degradation in DUT software.

Procedure: Record per-process memory usage at T=0, T=1h, T=6h, T=12h, T=24h. Compute linear regression slope of memory usage over time. A slope exceeding 1 MB/hour for any process indicates a potential memory leak and is reported; this slope is a reporting trigger for investigation, not a pass/fail criterion. Also monitor forwarding-plane counter wraparounds and hardware table occupancy trends.

13. Reporting Format

Per the BMWG charter, the definition of acceptance criteria or performance requirements is explicitly outside the scope of this Working Group. This methodology defines what is measured and how it is reported; it does not set minimum acceptable values, certification, or pass/fail criteria. Any deployment-specific performance objectives are outside the scope of this document.

Results from collective communication benchmarks (Section 9) MUST be reported per the reporting requirements stated in the BusBW definition of [TERMINOLOGY].

Test reports include the following sections:

  1. DUT Identification: Complete parameters from Section 3.3 for all fabric components.

  2. Test Topology: Diagram and description per Section 3.1, including physical cabling.

  3. Test Configuration: All DUT configuration parameters: QoS policies (ECN thresholds, PFC headroom, DCQCN parameters), load balancing mode, buffer allocation, and vendor-specific tuning.

  4. Host Configuration: Complete host stack description per Section 3.3 including NIC firmware, driver, collective library version, and any tuning. For UET tests, additionally report: UEC compliance profile, libfabric provider version, NIC UEC firmware version, and enabled optional features (LLR, Packet Trimming, Packet Rate Improvement (PRI), CBFC).

  5. Test Results: For each test from Section 5 through Section 12, provide specified tables, graphs, and statistical summaries. For Section 6 tests, results include side-by-side UET vs. RoCEv2 comparison data on the identical DUT fabric.

  6. Anomalies: Any deviations from specified procedures, test failures, or unexpected behaviors are documented.

  7. Repeatability Statement: Report iteration count and coefficient of variation (std deviation / mean) for each test's primary metric. A CV of 5% is an illustrative reference point for typical run-to-run variation; per the charter disclaimer above, this document does not set a required or minimum threshold for test validity.

  8. Comparability Statement: When a report compares two or more fabrics, tabulate the comparability set of Section 3.2 for each result and state explicitly which parameters differ. Reports comparing fabrics of different topology class additionally provide the structural descriptors listed in that section. Reports that present results from a single fabric do not require this section.

14. Security Considerations

This document defines benchmarking methodology for controlled laboratory environments and does not specify any protocol mechanism. It therefore introduces no new protocol-level security considerations beyond those of the underlying technologies it references. The considerations below follow the BMWG convention established in [RFC8238] and align with the companion terminology document [TERMINOLOGY].

Benchmarking activities as described in this document are limited to technology characterization of AI training fabrics using controlled stimuli in a laboratory environment, with dedicated address space and the constraints specified herein.

The benchmarking network topology will be an independent test setup and MUST NOT be connected to devices that may forward the test traffic into a production network or misroute traffic to the test management network. This isolation requirement is particularly important for AI fabric benchmarking because the hop-by-hop flow-control mechanisms referenced in this document (PFC, CBFC) propagate backpressure toward traffic sources and can extend the blast radius of a misconfigured test beyond the immediate DUT; DCQCN reduces, but does not eliminate, reliance on these mechanisms.

Benchmarking is performed on a "black-box" basis, relying solely on measurements observable external to the DUT as defined in [TERMINOLOGY].

Special capabilities SHOULD NOT exist in the DUT specifically for benchmarking purposes. Any implications for network security arising from the DUT SHOULD be identical in the lab and in production networks. In particular, RDMA memory-region permissions are properties of the deployed configuration, not of the benchmarking methodology, and SHOULD reflect production posture during testing.

Per [RFC6815], the tests defined herein MUST NOT be performed on production networks. The use of dedicated test IP address ranges per [RFC2544] Appendix C (198.18.0.0/15 for IPv4; 2001:db8::/32 per [RFC3849] for IPv6) is RECOMMENDED to prevent accidental interaction with production infrastructure.

The following considerations are specific to the methodology defined in this document:

15. IANA Considerations

This document has no IANA actions.

16. References

16.1. Normative References

[RFC1242]
Bradner, S., "Benchmarking Terminology for Network Interconnection Devices", RFC 1242, DOI 10.17487/RFC1242, , <https://www.rfc-editor.org/rfc/rfc1242>.
[RFC2119]
Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, DOI 10.17487/RFC2119, , <https://www.rfc-editor.org/rfc/rfc2119>.
[RFC2544]
Bradner, S. and J. McQuaid, "Benchmarking Methodology for Network Interconnect Devices", RFC 2544, DOI 10.17487/RFC2544, , <https://www.rfc-editor.org/rfc/rfc2544>.
[RFC6815]
Bradner, S., Dubray, K., McQuaid, J., and A. Morton, "Applicability Statement for RFC 2544: Use on Production Networks Considered Harmful", RFC 6815, DOI 10.17487/RFC6815, , <https://www.rfc-editor.org/rfc/rfc6815>.
[RFC8174]
Leiba, B., "Ambiguity of Uppercase vs Lowercase in RFC 2119 Key Words", BCP 14, RFC 8174, DOI 10.17487/RFC8174, , <https://www.rfc-editor.org/rfc/rfc8174>.
[RFC8238]
Avramov, L. and J. Rapp, "Data Center Benchmarking Terminology", RFC 8238, DOI 10.17487/RFC8238, , <https://www.rfc-editor.org/rfc/rfc8238>.
[RFC8239]
Avramov, L. and J. Rapp, "Data Center Benchmarking Methodology", RFC 8239, DOI 10.17487/RFC8239, , <https://www.rfc-editor.org/rfc/rfc8239>.
[RFC9004]
Morton, A., "Updates for the Back-to-Back Frame Benchmark in RFC 2544", RFC 9004, DOI 10.17487/RFC9004, , <https://www.rfc-editor.org/rfc/rfc9004>.
[TERMINOLOGY]
Calabria, F., Pignataro, C., Wu, Q., Fioccola, G., and S. Reddy, "Benchmarking Terminology for AI Network Fabrics", Work in Progress, Internet-Draft, draft-calabria-bmwg-ai-fabric-terminology-04, , <https://datatracker.ietf.org/doc/html/draft-calabria-bmwg-ai-fabric-terminology-04>.
[UEC-1.0]
Ultra Ethernet Consortium, "Ultra Ethernet Transport (UET) Specification 1.0", , <https://ultraethernet.org>.

16.2. Informative References

[DCQCN-PAPER]
Zhu, Y., Eran, H., Firestone, D., Guo, C., Lipshteyn, M., Liron, Y., Padhye, J., Raindel, S., Yahia, M. H., and M. Zhang, "Congestion Control for Large-Scale RDMA Deployments", ACM SIGCOMM pp. 523-536, DOI 10.1145/2785956.2787484, , <https://doi.org/10.1145/2785956.2787484>.
[EVPN-BENCH]
Jacob, S. and K. Tiruveedhula, "Benchmarking Methodology for EVPN and PBB-EVPN", Work in Progress, Internet-Draft, draft-ietf-bmwg-evpntest-11, , <https://datatracker.ietf.org/doc/html/draft-ietf-bmwg-evpntest-11>.
[INFERENCE-BENCH]
Calabria, F., Pignataro, C., Wu, Q., Fioccola, G., and S. Reddy, "Benchmarking Methodology for AI Inference Serving Network Fabrics", Work in Progress, Internet-Draft, draft-calabria-bmwg-ai-fabric-inference-bench-03, , <https://datatracker.ietf.org/doc/html/draft-calabria-bmwg-ai-fabric-inference-bench-03>.
[LIBFABRIC]
OpenFabrics Interfaces Working Group, "libfabric: Open Fabric Interfaces", n.d., <https://ofiwg.github.io/libfabric/>.
[LLM-BENCH]
Gaikwad, et al, "Benchmarking Methodology for Large Language Model Serving", Work in Progress, Internet-Draft, draft-gaikwad-llm-benchmarking-methodology-00, , <https://datatracker.ietf.org/doc/html/draft-gaikwad-llm-benchmarking-methodology-00>.
[META-ROCE]
Gangidi, A., Miao, R., Zheng, S., Bondu, S. J., Goes, G., Morsy, H., Puri, R., Riftadi, M., Shetty, A. J., Yang, J., Zhang, S., Fernandez, M. J., Gandham, S., and H. Zeng, "RDMA over Ethernet for Distributed Training at Meta Scale", ACM SIGCOMM '24 Sydney, NSW, Australia, DOI 10.1145/3651890.3672233, , <https://doi.org/10.1145/3651890.3672233>.
[MLPERF]
MLCommons, "MLPerf Training Benchmark Suite", n.d., <https://mlcommons.org>.
[RFC3849]
Huston, G., Lord, A., and P. Smith, "IPv6 Address Prefix Reserved for Documentation", RFC 3849, DOI 10.17487/RFC3849, , <https://www.rfc-editor.org/rfc/rfc3849>.

Appendix A. KPI-to-Test Mapping Summary

Table 15: KPI-to-Test Mapping Summary
KPI Test Section Measurement Method Reporting Unit
Throughput Rate Section 5.1 Binary search, zero-loss Tbps, % line rate
Latency (P99) Section 5.2 Tagged frame, loaded / unloaded us
Burst Absorption Section 5.3 Max burst without loss frames, bytes
ECN Accuracy Section 7.1 Queue depth vs. marking threshold deviation %
PFC Behavior Section 7.2 Incast sweep N=2..64 PAUSE events/sec, duration
DCQCN Convergence Section 7.3 Rate stabilization after onset us
PFC Deadlock Section 7.4 Cyclic adversarial traffic observed/reported, watchdog events
ECMP Imbalance Section 8.1 MMR, JFI per QP count dimensionless ratios
DLB Efficacy Section 8.2 Throughput delta vs. ECMP %, out-of-order rate
Spray Efficacy Section 8.3 JFI, retransmission rate dimensionless, retx/sec
AllReduce BusBW Section 9.1 CCL benchmark Gbps per accelerator
AllToAll JCT Section 9.2 CCL benchmark seconds per iteration
AllGather BusBW Section 9.3 CCL benchmark Gbps per accelerator
Synthetic JCT Ratio Section 10.1 Measured / Roofline dimensionless
MLPerf JCT Section 10.2 Time-to-train minutes, tokens/sec
Multi-Tenant Impact Section 10.3 Contention / Baseline JCT interference factor
Scale Limit Section 11.1 Max N with JCT Ratio characterized accelerator count
Failover Time Section 11.2 Loss duration on link fail us
24h Stability Section 12.1 JCT Ratio std deviation dimensionless
UET Throughput (RUD) Section 6.1 Binary search per transport service Gbps, % line rate
UET First-Packet Latency Section 6.2 PDC establish + first data us
UET Spray Efficacy Section 6.3 JFI/MMR under RUD spray dimensionless, OOO rate
UET PFC-Free Loss Rate Section 6.4 Incast without PFC enabled %, retx overhead
LLR Retry Latency Section 6.5 Per-hop error recovery time nanoseconds
Packet Trimming Savings Section 6.5 BW saved during congestion % bandwidth
CBFC vs PFC HOL Blocking Section 6.5 Head-of-line blocking duration us
UET Collective BusBW Section 6.6 AllReduce/AllToAll over UET Gbps per accelerator
PDC Establishment Rate Section 6.7 Sustained PDC creation rate PDCs/second
Max Concurrent PDCs Section 6.7 Scale limit per NIC count

Appendix B. Indicative Reference Values (Non-Normative)

This appendix provides indicative reference values for the KPIs defined in Section 4. The values reflect current industry observations for distributed AI training workloads as of 2025-2026. These values are NON-NORMATIVE and do not constitute benchmarking acceptance criteria or performance requirements. Per the BMWG charter, the definition of acceptance criteria or performance requirements is explicitly outside the scope of this Working Group. Implementers may use these values as contextual references when interpreting results; they MUST NOT be used as pass/fail criteria in vendor evaluations. Deployment-specific targets will vary by topology, accelerator architecture, collective library, and operator requirements.

Table 16: Indicative Reference Values for Distributed AI Training Fabrics (Non-Normative)
KPI Indicative Reference
JCT Ratio ≤ 1.05 (≤ 1.15 acceptable)
BusBW ≥ 90% of NIC line rate (intra-pod)
Aggregate Throughput ≥ 95% of bisection BW
Packet Drop Rate 0 ppm wire-level loss (lossless RoCEv2 profiles); 0 ppm application-visible loss (UET; see the Zero Packet Loss definition in [TERMINOLOGY])

Appendix C. ASIC Feature Categories (Informational)

This appendix identifies ASIC feature categories relevant to AI fabric performance. Implementers document which categories are present and enabled on the DUT. Specific vendor names are intentionally omitted.

Table 17: ASIC Feature Categories
Feature Category Sub-types Relevance to AI Fabric What to Report
Aggregate Switching BW ASIC-level capacity Cluster scale, bisection BW Total Tbps; per-port speed (400/800GbE)
Buffer Architecture Shared, VOQ, Cut-through Microburst absorption, PFC behavior, lossless operation Buffer type; total bytes; shared vs. dedicated split; per-port/queue allocation
Packet Distribution Per-flow, Per-packet, Flowlet ECMP load balancing quality and reordering risk Supported granularities; in-fabric reorder buffer (yes/no)
Congestion Control ECN marking, PFC, DCQCN DCQCN convergence and lossless behavior ECN granularity (port/queue/VOQ); PFC priorities; DCQCN parameter range
Adaptive Routing Flowlet, ECMP, Spray, Topology-aware Load balancing quality under collective patterns Algorithm type; flowlet gap timer range; topology-aware support
Telemetry Per-port, Per-queue, Per-flow Required for KPI measurement during benchmarking Monitoring granularity; streaming interval; INT support
Cluster Scale Support 2-tier, 3-tier Applicable topology scales Max cluster size per topology; ASIC count

All values are reported based on vendor documentation or measured capability. Additional DUT capabilities affecting benchmark results are also documented.

Appendix D. RoCEv2 Test Frame Format

Table 18: RoCEv2 Test Frame Format
Offset Field Size Value / Description
00 Ethernet Dst MAC 6B DUT next-hop MAC
06 Ethernet Src MAC 6B Test equipment MAC
12 EtherType / TPID 2B 0x0800 (IPv4) when untagged; 0x8100 (Tag Protocol Identifier — TPID) when 802.1Q-tagged
14 802.1Q Tag (optional) 4B When tagged: Tag Control Information (TCI: Priority Code Point (PCP)=3 for RoCEv2 priority, VLAN Identifier (VID)) followed by inner EtherType 0x0800. Omit this row entirely when untagged and shift subsequent offsets back by 4B
18 IPv4 Header 20B DSCP=26 (AF31, Assured Forwarding class 3, drop precedence 1), ECN=ECT(0) (ECN-Capable Transport), Proto=17 (UDP)
38 UDP Header 8B DstPort=4791 (RoCEv2), SrcPort=var
46 BTH (Base Transport Header) 12B OpCode, DstQP, PSN, P_Key
58 RDMA Extended Transport Header (RETH; if Write) 16B Virtual Address (VA), R_Key, Direct Memory Access (DMA) Length
74 Payload var Test data (incrementing octets)
var ICRC 4B Invariant CRC
var+4 FCS 4B Ethernet Frame Check Sequence

Appendix E. UET (Ultra Ethernet Transport) Frame Format

UET runs over UDP/IP using UDP destination port 4793, IANA-assigned to Ultra Ethernet Transport.

Unlike the RoCEv2 frame format above, this appendix does not specify a byte-accurate UET header layout. UET headers are layered and variable-length, and their wire formats are defined normatively by the UEC; test equipment implementations MUST follow [UEC-1.0] Section 4 for all wire-format details. The figure below is explicitly schematic: it shows only the layering and on-wire ordering of the protocol components that test equipment generates and parses.

+-------------------------------------------------------------+
| Ethernet Header: Dst MAC, Src MAC, EtherType                |
|   (optional 802.1Q tag: PCP=3 for UET priority class, VID)  |
+-------------------------------------------------------------+
| IPv4 Header: DSCP=26 (AF31), ECN=ECT(0), Proto=17 (UDP)     |
+-------------------------------------------------------------+
| UDP Header: DstPort=4793 (UET);                             |
|   SrcPort carries the Entropy Value, varied per packet      |
+-------------------------------------------------------------+
| PDS (Packet Delivery Sublayer) header(s):                   |
|   packet type, PDC identifiers, sequence number,            |
|   ack/credit and congestion-control state                   |
+-------------------------------------------------------------+
| SES (Semantic Sublayer) header(s):                          |
|   operation (Write/Read/Send/Atomic), addressing,           |
|   message identification                                    |
+-------------------------------------------------------------+
| Payload                                                     |
+-------------------------------------------------------------+
| ICRC (4B) | FCS (4B)                                        |
+-------------------------------------------------------------+
Figure 7: UET Frame Layering (Schematic Only; Byte Layouts Per UEC 1.0 Section 4)

Layering notes:

E.1. Key Differences from RoCEv2

Table 19: RoCEv2 vs. UET Comparison
Aspect RoCEv2 UET Notes
UDP Dst Port 4791 (IANA-assigned) 4793 (IANA-assigned) Distinct transports, both over UDP/IP
Transport Endpoint QP Number (24b) PDC ID PDC state is established in-band with the first packet
Entropy Source UDP src port, typically fixed per QP/connection UDP src port, varied per packet Per-packet spraying uses existing ECMP hashing
Congestion Signalling ECN bits; CNP-based feedback (e.g., DCQCN) ECN plus UET congestion-control state carried in PDS Sender- and receiver-based CC per [UEC-1.0]
Ordering Guarantee Always in-order (RC) Per-service (ROD/RUD/RUDI/UUD) RUD/RUDI allow out-of-order delivery
Header Structure Fixed BTH plus per-opcode extension headers Layered, variable-length (PDS, SES) Per [UEC-1.0] Section 4

  1. UDP Destination Port: UET uses port 4793 vs. RoCEv2 port 4791.

  2. Entropy Value: Carried in the UDP source port field and varied per packet for ECMP path selection. Test equipment varies the source port to achieve uniform path distribution.

  3. Transport Service Indicator: Header encodes transport service (ROD/RUD/RUDI/UUD). Tests set this to match the service being benchmarked.

  4. PDC Identifier: In-band-established PDC ID replaces RoCEv2's Destination QP. Test equipment tracks PDC lifecycle for accurate measurement.

  5. Layered Sub-Headers: UET uses four sub-layers (SES, PDS, CMS, TSS) with variable-length headers. Implementations MUST follow [UEC-1.0] Section 4 for wire format details.

  6. Optional Feature Headers: When the LLR or PRI link-layer features, or the network-layer Packet Trimming feature, are enabled, additional or modified framing may be present. Test equipment is configured to recognize and parse these.

Acknowledgments

This work has benefited from the discussions that occurred during the joint IPPM and BMWG meeting and on the BMWG mailing list. Thanks to Carsten Rossenhoevel and Mohamed Boucadair for valuable review and comments. Thanks to Andrew Yourtchenko for a thorough review of the document set. Thanks to Niangen Ye for the review comments on Fabric-Visible Data Volume provenance and on forwarding-work accounting, which prompted the byte-counting rule stated in Section 3.2.5.

Authors' Addresses

Fernando Calabria
Cisco
United States
Carlos Pignataro
Blue Fern Consulting
United States
Qin Wu
Huawei
China
Giuseppe Fioccola
Huawei
Italy
Sowjanya Reddy
Apple
United States