> 1. Does CSIT have any Broadcom bnxt PMD results, even for lower speeds?
No. > 2. Are there known issues or tuning recommendations for bnxt PMD at 400G? Not really, but there has been a thread [6] about a configuration issue, not sure how relevant. > 3. Would you be able to share the startup.conf and DPDK PMD settings used in the 3na-spr testbed? When looking at a test log, search for “Apply startup configuration on all VPP DUTs”. Examples: [7] [8]. > 4. Any known VPP graph node optimizations that help with 400G workloads? Native build should do the best optimizations already, but there might have been bugs (affecting only some nodes) preventing that optimum. Most of such performance bugs were minor, except this [9] one. The only universal recommendation is to use as few VPP workers as needed for the workload, because hardware prefetchers have hard time guessing which buffer will be accessed by which worker. In the past, we also had bad experiences with CPU stalls when changing frequency automatically (power saving or overheat prevention), but I hear it is less painful with newer CPUs. Vratko. [6] https://lists.fd.io/g/vpp-dev/topic/114778656?msg=26276#msg26276 [7] https://logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-report-coverage-2606-3na-spr/28635430896/log.html.gz#s1-s1-s1-s1-s2-t5-k2-k5-k1-k1-k1-k1 [8] https://logs.fd.io/vex-yul-rot-jenkins-1/csit-vpp-perf-report-coverage-2606-3na-spr/28566962521/log.html.gz#s1-s1-s1-s1-s8-t4-k2-k5-k1-k1-k1-k1 [9] https://github.com/FDio/vpp/issues/3579 From: [email protected] <[email protected]> On Behalf Of Dinesh via lists.fd.io Sent: Monday, 31 August, 2026 17:05 To: [email protected] Subject: Re: [vpp-dev] Query on 400-gig throughput Thanks for the detailed response. Our specific Setup: - VPP 25.02 with DPDK 24.11 (self-compiled) - Broadcom BCM5760X 400G NIC (device ID 14e4:1760) - Ubuntu 22.04 - [24] CPU cores, [128] GB RAM Questions: 1. Does CSIT have any Broadcom bnxt PMD results, even for lower speeds? 2. Are there known issues or tuning recommendations for bnxt PMD at 400G? 3. Would you be able to share the startup.conf and DPDK PMD settings used in the 3na-spr testbed? This would help us set a baseline. 4. Any known VPP graph node optimizations that help with 400G workloads? Would appreciate any pointers to Broadcom-specific tuning or benchmarks. Thanks On 31/08/26 17:45, Vratko Polak -X (vrpolak - PANTHEON TECHNOLOGIES at Cisco) via lists.fd.io wrote: The limit depends on NICs and ports available for the traffic generator machine. In CSIT, most testbeds have just one NIC per tested type, so even with bi-directional traffic we are limited on NIC despite using two ports. The proof is in one special testbed 3na-spr where we have multiple NICs of the same type (CX7, 4 NICs but two with only one port cabled). For the usual "one port per direction" tests, using ports on different NIC, we can achieve [1] ~400 Gbps there with large enough packets, and it also allows us to run "three ports per direction" tests achieving [2] ~800 Gbps. The topology file is here [3] and direct link to telemetry of randomly selected tests are [4] and [5]. Those are coverage tests executed only one per release, as results from single-physical-core tests are more stable for regression detection purposes. But let me know if you are interested in some specific one-off results I could produce.
-=-=-=-=-=-=-=-=-=-=-=- Links: You receive all messages sent to this group. View/Reply Online (#27159): https://lists.fd.io/g/vpp-dev/message/27159 Mute This Topic: https://lists.fd.io/mt/121010181/21656 Group Owner: [email protected] Unsubscribe: https://lists.fd.io/g/vpp-dev/leave/14379924/21656/631435203/xyzzy [[email protected]] -=-=-=-=-=-=-=-=-=-=-=-
