For ConnecX-6Dx and older HW with default kernel flow rules Rx hops per packet was about 2-3, Tx hops per packet was about 1-2. On BlueField 3 DPU without Open vSwitch or anything, DOCA Flow Tune from DOCA 3.2.1 LTS shows Rx hops per packet around 40-50, Tx hops per packet around 12. This suggests that hops meaning is different for BlueField 3 and older HW. Is it documented anywhere? For older HW, there was a notion of optimal Rx/Tx hops per packet depending on packets per second and chip frequency. It there a similar limit for BlueField 3? DOCA Flow Tune description of âRx/Tx hopsâ is too terse.
Hi @dmitry.kozliuk,
On ConnectXâ6Dx and earlier NICs, the default kernel steering pipeline is relatively shallow, so DOCA Flow Tune typically reports around 2â3 Rx hops and 1â2 Tx hops per packet.
On BlueFieldâ3, the DOCA Flow / steering pipeline is more hierarchical (more internal tables, anchors, possible recirculation), so a single packet can legitimately traverse more internal match/action stages. As a result, higher Rx/Tx hops per packet values on BF3 (e.g. tens of hops) do not map 1:1 to the â2â3 / 1â2 hopsâ you may have seen on older NIC generations. The âRx/Tx hops per packetâ counters in DOCA Flow Tune are counting internal steering hops in the RX/TX domains, not network hops â
Unlike older ConnectX NICs, there is no published formula or âoptimal hops vs pps vs frequencyâ guideline for BlueFieldâ3. These counters are intended to compare different pipelines on the same BF3 platform (for example, before/after a steering change), together with other KPIs such as ICM cacheâmiss per packet and Rx/Tx packet rates, rather than to enforce a fixed absolute hop limit.
In practice, as long as your throughput/latency targets are met and ICM/cacheârelated counters look healthy, higher hops values on BF3 alone do not indicate a problem â but any tuning should always be validated in your own traffic patterns and environment. Of course, if you see any serious performance degradation or behavior that clearly does not match expectations, please open an Enterprise Support ticket so the NVIDIA support team can review your specific setup and traces in detail.
xyin