--- title: "AI Fabric and RoCEv2 Switch Testing: PFC, ECN, Incast, ECMP, Failover | TEST4NET" description: "Independent testing of AI and storage fabric switches in CloudLAB with RoCEv2 traffic on 100G ports: PFC and ECN under incast, buffer headroom, ECMP balance, latency per flow, link failover. Test plan in 1–3 days." image: https://test4net.com/hubfs/og/solutions-ai-fabric-rocev2.png --- [Skip to content](https://test4net.com/solutions/ai-fabric-rocev2#main-content) [![TEST4NET](https://test4net.com/hubfs/test4net-logo-fixed-2.svg)](https://test4net.com/?hsLang=en) - Services - Products - Resources - Company [News](https://test4net.com/news?hsLang=en) [Login](https://lab.test4net.com) [Support](https://test4net.com/support?hsLang=en) What we test By technology Industries [Load & Performance Testing Throughput, sessions and latency up to and past the rated limit](https://test4net.com/services/load-performance-testing?hsLang=en) [Functional & Conformance Testing Features against the spec, RFC compliance with evidence](https://test4net.com/services/functional-conformance-testing?hsLang=en) [Interoperability Testing Multi-vendor behaviour, before it reaches production](https://test4net.com/services/interoperability-testing?hsLang=en) [Security & Resilience Testing NGFW, IPS and VPN efficacy and performance with threats on the wire](https://test4net.com/services/security-resilience-testing?hsLang=en) [Test Plan & Methodology Design What to test, how, and what counts as a pass](https://test4net.com/services/test-plan-design?hsLang=en) [Test Automation & CI/CD Regression suites wired into your release pipeline](https://test4net.com/services/test-automation?hsLang=en) [Training Programs Hands-on courses for your test engineers](https://test4net.com/services#training) [Equipment Repair Diagnostics and repair of test equipment](https://test4net.com/services#repair) [All services](https://test4net.com/services?hsLang=en) [Routing & Switching RFC 2544 and RFC 2889 on real ports, 1G to 100G](https://test4net.com/solutions/routing-switching?hsLang=en) [Firewalls & NGFW RFC 9411 throughput with inspection switched on](https://test4net.com/solutions/firewalls-ngfw?hsLang=en) [SD-WAN & WAN Path selection and QoS under emulated WAN conditions](https://test4net.com/solutions/sd-wan?hsLang=en) [Cloud & NFV Virtual appliances and cloud services under load](https://test4net.com/solutions/cloud-nfv?hsLang=en) [5G & Mobile Core Transport and core functions under subscriber traffic](https://test4net.com/solutions/5g-mobile-core?hsLang=en) [Wi-Fi Wi-Fi 6 clients and access points, Wi-Fi 7 in progress](https://test4net.com/solutions/wi-fi?hsLang=en) [Storage & Infrastructure Storage networks and servers under realistic I/O](https://test4net.com/solutions/storage?hsLang=en) [AI / Switching Fabric (RoCEv2) Lossless Ethernet for GPU clusters: PFC, ECN, fabric load](https://test4net.com/solutions/ai-fabric-rocev2?hsLang=en) [All technologies](https://test4net.com/services#coverage) [Telecom Operators Pre-deployment acceptance and vendor comparison](https://test4net.com/solutions/industries#telecom) [Network & Security Vendors Independent validation of your product claims](https://test4net.com/solutions/industries#vendors) [System Integrators Proof for the customer before the rollout](https://test4net.com/solutions/industries#integrators) [Enterprise & Banks Upgrades and new platforms checked before production](https://test4net.com/solutions/industries#enterprise) [All industries](https://test4net.com/solutions/industries?hsLang=en) Not sure where to start? Test plan and quote in 1–3 business days Tell us what you run and what worries you. We propose the tests, tools and timeline. [Request a test plan](https://test4net.com/contacts?hsLang=en) CloudLAB [How It Works Pick a scenario, run it in the live lab, get the report](https://test4net.com/cloudlab#how) [Equipment & Capabilities L2-3 to 100G, L4-7 past 200G, Wi-Fi, WAN emulation](https://test4net.com/cloudlab#equipment) [Access Plans Hourly sessions, subscriptions, dedicated racks](https://test4net.com/cloudlab/plans?hsLang=en) [Hosted DUT Testing Ship the box or upload the VM, we run the tests](https://test4net.com/cloudlab/hosted-dut?hsLang=en) [Open the CloudLAB portal](https://lab.test4net.com) Our tools [Traffic Generator L4-7 and security traffic past 200 Gbit/s, app mix library](https://test4net.com/products/traffic-generator?hsLang=en) [WANemu Delay, jitter, loss and bandwidth shaping per link](https://test4net.com/products/wanemu?hsLang=en) [Book a demo](https://test4net.com/contacts?hsLang=en) New: our own test tools Our own tools, in beta this autumn Built in our lab, priced per gigabit. Join the beta list and see them first. [Join the beta list](https://test4net.com/products#lead-form) Read Technologies / Protocols Methodologies [Blog Case studies, test insights and lab notes](https://test4net.com/blog?hsLang=en) [Industry News What is changing in network testing and infrastructure](https://test4net.com/news?hsLang=en) [FAQ How the lab, sessions and reports work](https://test4net.com/faq?hsLang=en) [All resources](https://test4net.com/resources?hsLang=en) [BGP Route exchange, policy and convergence](https://test4net.com/blog/how-we-test-bgp?hsLang=en) [OSPF & IS-IS IGP adjacency, flooding and SPF timing](https://test4net.com/blog/how-we-test-ospf-is-is?hsLang=en) [MPLS & SR LDP, RSVP‑TE and SR‑MPLS label paths](https://test4net.com/blog/how-we-test-mpls-and-segment-routing?hsLang=en) [EVPN / VXLAN Overlay fabrics, MAC/IP routes, multi-homing](https://test4net.com/blog/how-we-test-evpn-vxlan?hsLang=en) [IPsec & IKEv2 Tunnel setup, rekeying and throughput](https://test4net.com/blog/how-we-test-ipsec-and-ikev2?hsLang=en) [TLS / HTTPS Handshakes, cipher suites and inspection](https://test4net.com/blog/how-we-test-tls-and-https?hsLang=en) [RoCEv2 Lossless Ethernet, PFC and ECN for AI fabrics](https://test4net.com/blog/how-we-test-rocev2?hsLang=en) [Wi-Fi 6 / 6E / 7 802.11ax/be, OFDMA, MLO and roaming](https://test4net.com/blog/how-we-test-wi-fi-6-6e-7?hsLang=en) [All protocols](https://test4net.com/blog/tag/protocols?hsLang=en) [RFC 2544 Throughput, latency, frame loss and back-to-back](https://test4net.com/methodologies/rfc-2544?hsLang=en) [RFC 2889 LAN switch benchmarking: forwarding, MAC learning, congestion](https://test4net.com/methodologies/rfc-2889?hsLang=en) [RFC 9411 / RFC 3511 Firewall and NGFW performance benchmarking](https://test4net.com/methodologies/rfc-9411-rfc-3511?hsLang=en) [RFC 6349 TCP throughput and buffer delay testing](https://test4net.com/methodologies/rfc-6349?hsLang=en) [RFC 5180 IPv6 benchmarking methodology](https://test4net.com/methodologies/rfc-5180?hsLang=en) [ITU-T Y.1564 Ethernet service activation: CIR, EIR, KPIs per service](https://test4net.com/methodologies/y-1564?hsLang=en) [ITU-T Y.1731 Ethernet OAM: delay, jitter, loss and SLA monitoring](https://test4net.com/methodologies/y-1731?hsLang=en) [TR-398 Wi-Fi in-premises performance test plan](https://test4net.com/methodologies/tr-398?hsLang=en) [All methodologies](https://test4net.com/methodologies?hsLang=en) Read [Case studies, test insights and lab notes Blog](https://test4net.com/blog?hsLang=en) [What is changing in network testing and infrastructure Industry News](https://test4net.com/news?hsLang=en) [How the lab, sessions and reports work FAQ](https://test4net.com/faq?hsLang=en) [About Us Independent lab in Armenia, team and history](https://test4net.com/about?hsLang=en) [Why TEST4NET What you get that vendor slides never show](https://test4net.com/why-test4net?hsLang=en) [Lab & Certifications Bench capacity, methods and the standards we test against](https://test4net.com/lab?hsLang=en) [Partners Test tool vendors and IT partners we work with](https://test4net.com/partners?hsLang=en) [Contacts Talk to an engineer, not a sales bot](https://test4net.com/contacts?hsLang=en) [Support portal](https://test4net.com/support?hsLang=en) - Services What we test [Load & Performance Testing](https://test4net.com/services/load-performance-testing?hsLang=en) [Functional & Conformance Testing](https://test4net.com/services/functional-conformance-testing?hsLang=en) [Interoperability Testing](https://test4net.com/services/interoperability-testing?hsLang=en) [Security & Resilience Testing](https://test4net.com/services/security-resilience-testing?hsLang=en) [Test Plan & Methodology Design](https://test4net.com/services/test-plan-design?hsLang=en) [Test Automation & CI/CD](https://test4net.com/services/test-automation?hsLang=en) [Training Programs](https://test4net.com/services#training) [Equipment Repair](https://test4net.com/services#repair) [All services](https://test4net.com/services?hsLang=en) By technology [Routing & Switching](https://test4net.com/solutions/routing-switching?hsLang=en) [Firewalls & NGFW](https://test4net.com/solutions/firewalls-ngfw?hsLang=en) [SD-WAN & WAN](https://test4net.com/solutions/sd-wan?hsLang=en) [Cloud & NFV](https://test4net.com/solutions/cloud-nfv?hsLang=en) [5G & Mobile Core](https://test4net.com/solutions/5g-mobile-core?hsLang=en) [Wi-Fi](https://test4net.com/solutions/wi-fi?hsLang=en) [Storage & Infrastructure](https://test4net.com/solutions/storage?hsLang=en) [AI / Switching Fabric (RoCEv2)](https://test4net.com/solutions/ai-fabric-rocev2?hsLang=en) [All technologies](https://test4net.com/services#coverage) Industries [Telecom Operators](https://test4net.com/solutions/industries#telecom) [Network & Security Vendors](https://test4net.com/solutions/industries#vendors) [System Integrators](https://test4net.com/solutions/industries#integrators) [Enterprise & Banks](https://test4net.com/solutions/industries#enterprise) [All industries](https://test4net.com/solutions/industries?hsLang=en) [Not sure where to start? Test plan and quote in 1–3 business days](https://test4net.com/contacts?hsLang=en) - Products CloudLAB [How It Works](https://test4net.com/cloudlab#how) [Equipment & Capabilities](https://test4net.com/cloudlab#equipment) [Access Plans](https://test4net.com/cloudlab/plans?hsLang=en) [Hosted DUT Testing](https://test4net.com/cloudlab/hosted-dut?hsLang=en) [Open the CloudLAB portal](https://lab.test4net.com) Our tools [Traffic Generator](https://test4net.com/products/traffic-generator?hsLang=en) [WANemu](https://test4net.com/products/wanemu?hsLang=en) [Book a demo](https://test4net.com/contacts?hsLang=en) [New: our own test tools Our own tools, in beta this autumn](https://test4net.com/products#lead-form) - Resources Read [Blog](https://test4net.com/blog?hsLang=en) [Industry News](https://test4net.com/news?hsLang=en) [FAQ](https://test4net.com/faq?hsLang=en) [All resources](https://test4net.com/resources?hsLang=en) Technologies / Protocols [BGP](https://test4net.com/blog/how-we-test-bgp?hsLang=en) [OSPF & IS-IS](https://test4net.com/blog/how-we-test-ospf-is-is?hsLang=en) [MPLS & SR](https://test4net.com/blog/how-we-test-mpls-and-segment-routing?hsLang=en) [EVPN / VXLAN](https://test4net.com/blog/how-we-test-evpn-vxlan?hsLang=en) [IPsec & IKEv2](https://test4net.com/blog/how-we-test-ipsec-and-ikev2?hsLang=en) [TLS / HTTPS](https://test4net.com/blog/how-we-test-tls-and-https?hsLang=en) [RoCEv2](https://test4net.com/blog/how-we-test-rocev2?hsLang=en) [Wi-Fi 6 / 6E / 7](https://test4net.com/blog/how-we-test-wi-fi-6-6e-7?hsLang=en) [All protocols](https://test4net.com/blog/tag/protocols?hsLang=en) Methodologies [RFC 2544](https://test4net.com/methodologies/rfc-2544?hsLang=en) [RFC 2889](https://test4net.com/methodologies/rfc-2889?hsLang=en) [RFC 9411 / RFC 3511](https://test4net.com/methodologies/rfc-9411-rfc-3511?hsLang=en) [RFC 6349](https://test4net.com/methodologies/rfc-6349?hsLang=en) [RFC 5180](https://test4net.com/methodologies/rfc-5180?hsLang=en) [ITU-T Y.1564](https://test4net.com/methodologies/y-1564?hsLang=en) [ITU-T Y.1731](https://test4net.com/methodologies/y-1731?hsLang=en) [TR-398](https://test4net.com/methodologies/tr-398?hsLang=en) [All methodologies](https://test4net.com/methodologies?hsLang=en) Read [Blog Case studies, test insights and lab notes](https://test4net.com/blog?hsLang=en) [Industry News What is changing in network testing and infrastructure](https://test4net.com/news?hsLang=en) [FAQ How the lab, sessions and reports work](https://test4net.com/faq?hsLang=en) - Company [About Us](https://test4net.com/about?hsLang=en) [Why TEST4NET](https://test4net.com/why-test4net?hsLang=en) [Lab & Certifications](https://test4net.com/lab?hsLang=en) [Partners](https://test4net.com/partners?hsLang=en) [Contacts](https://test4net.com/contacts?hsLang=en) [News](https://test4net.com/news?hsLang=en) [Login](https://lab.test4net.com) [Support](https://test4net.com/support?hsLang=en) [Services](https://test4net.com/services?hsLang=en) / AI fabric and RoCEv2 # AI fabric and RoCEv2: the lossless check before the GPUs are plugged in We test the leaf and spine switches of an AI or storage fabric on the CloudLAB bench with RoCEv2 traffic on 100G ports: PFC and ECN behaviour under incast, buffer headroom, ECMP balance across the uplinks, latency per flow at load, and a link failure with flows in flight. The scope follows the bench: what is on it today goes into the plan, nothing else. A run takes from 2 weeks and ends with a report that holds drops and pause time per queue at each incast ratio, p99 latency, link balance and failover time against your fabric design. Test plan and quote in 1–3 business days. [Request a test plan](https://test4net.com/solutions/ai-fabric-rocev2#lead-form) [See what we test](https://test4net.com/solutions/ai-fabric-rocev2#classes) **Queue depth under 8:1 incast** One leaf port, 100G, the same 2 ms window before and after tuning 0200400 QUEUE, KB 00.511.52 TIME, MS ECN min PFC pause drops Tuned config, day 7 Delivered config Delivered config Tuned config, day 7 - **100G ports**RoCEv2 endpoints emulated on the bench - **4 checks**PFC and ECN, incast, ECMP, failover - **From 2 weeks**for a run - **1–3 days**test plan and quote When the number is needed ## Three points in a fabric project where a vendor slide is not enough A lossless Ethernet fabric is a set of thresholds on every switch. Each point below asks whether the thresholds on the switches match the traffic the cluster will produce, and each gets its own scope. The cases stay the same, so the shortlist number, the acceptance number and the post-upgrade number compare. 1. 1 ### Shortlist Tender, 2–3 switch models Every candidate runs the same incast, balance and failover cases with the same RoCEv2 endpoints. Vendors see the plan before the first packet and the results after the last. Ranked table per KPI, buffer behaviour per model, 3–4 weeks for 3 models 2. 2 ### Acceptance Before the first training job The delivered leaf and spine with the production config: DSCP to queue mapping, PFC priority, ECN thresholds, buffer profile and hashing. Measured against your fabric design, before the GPUs arrive or before the first job. Signed report with pass and fail per case, config exports, from 2 weeks 3. 3 ### Change NOS upgrade or new thresholds The acceptance cases run again on the new release or the new buffer profile. What moved, by how much, and whether the fabric still clears the design. The cluster keeps running; we test the lab pair or a spare leaf. Diff per KPI against the last run, 3 business days What we test ## Four fabric pieces, and the numbers each one has to show The bench is the same, the traffic and the pass criteria are not. The table is the starting point of the plan; your fabric design and the job profile set the thresholds. | Piece under test | Numbers in the report | Traffic and config on the bench | Typical size | | --- | --- | --- | --- | | 01Leaf switch, one box | Drops and PFC pause per queue at 2:1 to 16:1 incast, ECN marks against thresholds, latency per flow | RoCEv2 flows from emulated endpoints next to TCP, your DSCP map, PFC, ECN and buffer profile | 32–64 ports, 100G | | 02Leaf and spine pod | ECMP balance across uplinks, drops under cross-leaf incast, rehash time on uplink loss | Long and short flows between leaves through the spine, one uplink taken down with traffic in flight | 2 leaves and 1–2 spines, 100G | | 03Storage fabric | Lossless storage queue next to TCP, pause propagation between ports, latency per I/O-sized flow | NVMe over RoCEv2 flows and a TCP mix on one leaf, storage priority as configured | 1 leaf, 25G and 100G | | 04Congestion control on the NIC | Rate reduction and recovery on ECN marks, fairness on a shared port, settle time after a burst | Your NIC behind the leaf, DCQCN or the vendor profile, agreed per setup when the bench has the card | 2–8 NICs, 100G | Any switch vendor and NOS. 100G ports with emulated RoCEv2 endpoints on the bench today; 400G and GPU-level collective runs are agreed per setup. Named test tools are picked for the setup and listed in the report. The bench around one leaf and spine pod Senders 8 RoCEv2 endpoints, 100G TCP background storage and management mix Spine Leaf-1Leaf-2 Leaf and spine under test DSCP, PFC, ECN, buffers Receiver port incast target, 1 port Counters, captures per queue, from the switch LINK DOWN ON PURPOSE PFC pauses · ECN marks drops · latency per flow Around the switch ## What sits on each side of the leaf while it is measured A lossless number only means something with the right things around the switch: the many-to-one traffic a collective produces, the thresholds your config sets, and the faults the fabric has to survive. All three are on the bench at once. 1. ### Incast on purpose Emulated RoCEv2 endpoints send into one receiver port at 2:1, 4:1, 8:1 and 16:1, with a TCP background on the neighbouring ports. That is what the last step of an all-reduce looks like to a leaf. 2. ### Your config, line by line DSCP to queue mapping, the PFC priority, ECN min and max, the buffer profile and the hash fields from your export. The report shows which line each drop or pause traces back to. 3. ### Faults with flows in flight One uplink down, a spine reload, a port flap on the receiver. Rehash time, drops and pause time during each event, timed and logged. Typical scenario ## A pod for 64 GPUs, a datasheet that said lossless, and 0.3% of packets that were not An integrator hands over the leaf and spine of a training pod for acceptance, two weeks before the GPUs. The datasheet says lossless Ethernet. With the delivered config an 8:1 incast drops 0.3% of RoCEv2 packets and p99 latency is 48 µs against a design of 20. The run log points at ECN switched off on the RoCE queue and PFC enabled on the wrong priority. The integrator and the vendor fix the thresholds and the buffer profile on day 7, the rerun gives 0 drops and 11 µs, ECMP imbalance across the uplinks goes from 38% to 6% after a hash change. The pod is accepted on a signed report, and the same cases run again after the next NOS release in 3 days. 0.3% → 0 drops at 8:1 incast, before and after tuning 48 → 11 µs p99 latency per flow at 90% load 11 days from the first packet to the signed report Composite scenario from projects of this type. Numbers are realistic and not from one client. p99 latency at 8:1 incast, µs Delivered config Tuned config, day 7 48 11 Design 20 0204060 above the design clears the design What you take away ## Four numbers that end the argument, and the table they live in The report is built for the meeting after the test: the vendor, the integrator and your platform team read the same table, with the pass band from the fabric design next to the measured value Each number comes with its evidence: generator statistics, switch counters per queue, captures and the config export of the run Lossless Drops and PFC pause time per queue at each incast ratio, and the ECN mark rate against the configured thresholds. Not the datasheet word. Latency p50 and p99 per flow at 50% and 90% load, with and without incast, so the fabric design has a latency budget per hop Balance Load per uplink under the configured hashing, the spread between the busiest and the quietest link, and the flow sizes that skew it Failover Milliseconds until flows settle on the remaining uplinks and how many packets were lost, with flows in flight, not on an idle pod ACC-AI-01 Acceptance, leaf and spine pod Rerun, day 7 ### Result table, 6 of 12 cases | KPI | Pass band | Measured | Result | | --- | --- | --- | --- | | Drops at 8:1 incast, RoCE queue | 0 | 0 | Pass | | PFC pause time per minute | ≤ 50 ms | 12 ms | Pass | | p99 latency at 90% load | ≤ 20 µs | 11 µs | Pass | | ECMP imbalance across 4 uplinks | ≤ 10% | 6% | Pass | | Uplink down, flows in flight | ≤ 100 ms | 70 ms | Pass | | ECN marks start | at 150 KB | from 180 KB | Pass with a note | Where it runs ## The CloudLAB switching bench, and the pages behind this one Switches ship to CloudLAB or we connect the bench to your lab, and you watch the run through the portal. The bench has 100G ports with emulated RoCEv2 endpoints, PFC and ECN today; it grows step by step and the plan lists only what is on it for your run. The frame-level base is RFC 2544 and RFC 2889; the lossless cases are written procedures you sign off. [How a load run goes, step by step](https://test4net.com/services/load-performance-testing?hsLang=en) [RFC 2544: throughput, latency, loss](https://test4net.com/methodologies/rfc-2544?hsLang=en) [RFC 2889: switch benchmarks](https://test4net.com/methodologies/rfc-2889?hsLang=en) [Get the test plan written first](https://test4net.com/services/test-plan-design?hsLang=en) ### Related reading - [BlogLossless Ethernet for AI fabrics: PFC, LLR or CBFC, and how to verify it](https://test4net.com/blog/lossless-ethernet-ai-fabrics-pfc-llr-cbfc?hsLang=en) - [BlogUltra Ethernet 1.1 vs RoCEv2: what changes in the transport and how to test it](https://test4net.com/blog/ultra-ethernet-1-1-vs-rocev2-ai-fabric-transport-testing?hsLang=en) - [BlogFlowlet balancing evens out AI fabric traffic where ECMP piles it up](https://test4net.com/blog/keysight-flowlet-ai-ecmp?hsLang=en) - [BlogMeta at OCP 2026: no fabric accepted without benchmarks](https://test4net.com/blog/meta-ocp-2026-ai-fabric-validation-benchmarks?hsLang=en) Before the brief ## What fabric buyers ask first Which switches can you test? Any vendor and any NOS: leaf and spine boxes with 25G and 100G ports, as one switch or as a small pod of 2 leaves and 1–2 spines. We need the models, the port plan and an export of the config; the rest is in the plan. Do we need GPUs on the bench? No. The RoCEv2 flows come from emulated endpoints on the bench, which is what lets us run a 16:1 incast on demand and repeat it after a change. A collective run on real GPUs is a different test and is agreed per setup, with your servers or ours written into the plan. Is it tested with our config or with vendor defaults? With your config: DSCP to queue mapping, PFC priority, ECN min and max, buffer profile and hash fields from an anonymised export. A vendor-default run can be added as a second baseline so you see what each threshold changes. What about 400G, 800G and Ultra Ethernet? The bench runs 100G ports today and grows step by step. A 400G leaf is agreed per setup and goes into the plan only when the bench has the ports for it; we say so before the quote, not after. Ultra Ethernet is on the blog and in our own development, not on the bench yet. Do you ship the switch back? Yes. Boxes go into the CloudLAB rack with KVM and console access and return after the run. For a fabric that is already racked we connect the bench to your lab over a dedicated link and test a spare leaf or the lab pair. How long does it take and what do we get? Test plan and quote in 1–3 business days after the brief. A run takes from 2 weeks for one leaf and 3–4 weeks for a shortlist of 3 models or a pod. The report holds the result table, queue counters, latency histograms, captures and config exports, and the rerun of fixed items is part of the plan. ## Send the switch models, the port plan and the config export. The plan comes back in 1–3 business days. No prices on the site: scope first, then the quote. The vendor and the integrator can sit in on the brief. [Request a test plan](https://test4net.com/solutions/ai-fabric-rocev2#lead-form) [How the load run goes](https://test4net.com/services/load-performance-testing?hsLang=en) Send the shortlist ## You have the method. Now get the numbers. Name the models and the KPIs that matter. An engineer replies within one business day with a test plan draft and a quote. Prefer to chat? [Message an engineer on WhatsApp](https://wa.me/37491612709) We use what you send only to answer you. See our [Privacy Policy](https://test4net.com/privacy-policy?hsLang=en). [Send the shortlist](https://test4net.com/solutions/ai-fabric-rocev2#lead-form) [![TEST4NET LLC](https://test4net.com/hubfs/test4net-logo-fixed.svg)](https://test4net.com?hsLang=en) Independent testing and validation for telecom, information security and cloud infrastructure – real insights, not assumptions [mailto:sales@test4net.com](mailto:sales@test4net.com) #### Navigate - [Home](https://test4net.com/?hsLang=en) - [Why TEST4NET](https://test4net.com/why-test4net?hsLang=en) - [About us](https://test4net.com/about?hsLang=en) - [Contact](https://test4net.com/contacts?hsLang=en) - [Blog](https://test4net.com/blog?hsLang=en) - [Industry News](https://test4net.com/news?hsLang=en) #### Get in touch Sales [sales@test4net.com](mailto:sales@test4net.com) Support [support.test4net.com](https://support.test4net.com) © 2026 TEST4NET LLC. All rights reserved. [Privacy Policy](https://test4net.com/privacy-policy?hsLang=en) [Legal](https://test4net.com/legal?hsLang=en) ```json { "@context" : "https://schema.org", "@id" : "https://test4net.com/#organization", "@type" : "Organization", "address" : { "@type" : "PostalAddress", "addressCountry" : "AM", "addressLocality" : "Yerevan", "postalCode" : "0033", "streetAddress" : "3 Hakob Hakobyan Street (Mergelyan Hub)" }, "alternateName" : "TEST4NET", "contactPoint" : { "@type" : "ContactPoint", "availableLanguage" : [ "en", "ru", "hy" ], "contactType" : "sales", "email" : "sales@test4net.com", "telephone" : "+374-60-88-88-99" }, "description" : "Independent, vendor-neutral lab for testing network and cybersecurity products — NGFW, WAF, IPS, DLP and SASE — and an authorized distributor of test, measurement and network infrastructure solutions.", "email" : "sales@test4net.com", "foundingDate" : "2024", "image" : "https://test4net.com/hubfs/og-t4n.png", "logo" : "https://test4net.com/hubfs/test4net-logo-fixed-2.svg", "name" : "TEST4NET LLC", "sameAs" : [ "https://www.linkedin.com/company/test4net/", "https://t.me/TEST4NET", "https://www.facebook.com/people/Test4net-LLC/61591470856077/" ], "telephone" : "+374-60-88-88-99", "url" : "https://test4net.com/" } ``` ```json { "@context" : "https://schema.org", "@id" : "https://test4net.com/#website", "@type" : "WebSite", "inLanguage" : "en", "name" : "TEST4NET", "potentialAction" : { "@type" : "SearchAction", "query-input" : "required name=search_term_string", "target" : { "@type" : "EntryPoint", "urlTemplate" : "https://test4net.com/hs-search-results?term={search_term_string}" } }, "publisher" : { "@id" : "https://test4net.com/#organization" }, "url" : "https://test4net.com/" } ``` ```json { "@context" : "https://schema.org", "@type" : "Service", "areaServed" : "Worldwide", "description" : "We test the leaf and spine switches of an AI or storage fabric on the CloudLAB bench with RoCEv2 traffic on 100G ports: PFC and ECN behaviour under incast, buffer headroom, ECMP balance across the uplinks, latency per flow at load, and a link failure with flows in flight. The scope follows the bench: what is on it today goes into the plan, nothing else. A run takes from 2 weeks and ends with a report that holds drops and pause time per queue at each incast ratio, p99 latency, link balance and failover time against your fabric design. Test plan and quote in 1–3 business days.", "isPartOf" : { "@type" : "WebPage", "url" : "https://test4net.com/services" }, "name" : "AI fabric and RoCEv2", "provider" : { "@type" : "Organization", "name" : "TEST4NET", "url" : "https://test4net.com/" }, "serviceType" : "AI and storage fabric switch testing with RoCEv2 traffic: PFC and ECN behaviour under incast, buffer headroom, ECMP balance, latency, link failover", "url" : "https://test4net.com/solutions/ai-fabric-rocev2" } ``` ```json { "@context" : "https://schema.org", "@type" : "BreadcrumbList", "itemListElement" : [ { "@type" : "ListItem", "item" : "https://test4net.com/services", "name" : "Services", "position" : 1 }, { "@type" : "ListItem", "item" : "https://test4net.com/solutions/ai-fabric-rocev2", "name" : "AI fabric and RoCEv2", "position" : 2 } ] } ``` ```json { "@context" : "https://schema.org", "@type" : "FAQPage", "mainEntity" : [ { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "Any vendor and any NOS: leaf and spine boxes with 25G and 100G ports, as one switch or as a small pod of 2 leaves and 1–2 spines. We need the models, the port plan and an export of the config; the rest is in the plan." }, "name" : "Which switches can you test?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "No. The RoCEv2 flows come from emulated endpoints on the bench, which is what lets us run a 16:1 incast on demand and repeat it after a change. A collective run on real GPUs is a different test and is agreed per setup, with your servers or ours written into the plan." }, "name" : "Do we need GPUs on the bench?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "With your config: DSCP to queue mapping, PFC priority, ECN min and max, buffer profile and hash fields from an anonymised export. A vendor-default run can be added as a second baseline so you see what each threshold changes." }, "name" : "Is it tested with our config or with vendor defaults?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "The bench runs 100G ports today and grows step by step. A 400G leaf is agreed per setup and goes into the plan only when the bench has the ports for it; we say so before the quote, not after. Ultra Ethernet is on the blog and in our own development, not on the bench yet." }, "name" : "What about 400G, 800G and Ultra Ethernet?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "Yes. Boxes go into the CloudLAB rack with KVM and console access and return after the run. For a fabric that is already racked we connect the bench to your lab over a dedicated link and test a spare leaf or the lab pair." }, "name" : "Do you ship the switch back?" }, { "@type" : "Question", "acceptedAnswer" : { "@type" : "Answer", "text" : "Test plan and quote in 1–3 business days after the brief. A run takes from 2 weeks for one leaf and 3–4 weeks for a shortlist of 3 models or a pod. The report holds the result table, queue counters, latency histograms, captures and config exports, and the rerun of fixed items is part of the plan." }, "name" : "How long does it take and what do we get?" } ] } ```