Is Direct Hardware Port Forwarding with line-rate Conversion possible with ConnectX-5 ?

We have a ConnectX-5 NIC with Port 0 connected to a 100 GbE link partner; and port 1 connected to a 1 GbE link partner. Fine. We can do everything one would expect you can do with the OS shoveling packets between ports. But we are doing latency-jitter-sensitive design and test. And we desire the capability to temporarily and exclusively make the ConnectX-5 a line-rate converting port forwarder. Of course a MLNX switch does this. But we wanted to ask if, maybe as a side effect of some other feature, this is possible? We don’t need any host-access or NIC function when in this (probably non-existing) ConnectX-5 operating mode. Thanks in advance. -Shep

Are you looking for configuration that will allow forward packets from one port to another without sending the packets to OS level? If yes, you probably searching for “Host Chaining” feature

Thank you @Aleksey Senin​ ! That’s exactly what we are looking for. Turns out in the meantime we got a Mellanox SN2010 switch, which is perhaps overkill - and a very nice switch and OS. But I was unaware of the term “Host Chaining” at the NIC. Terrific. Thanks! -Shep Siegel

Sorry @Aleksey Senin​ and Mellanox Team, We’re still stuck on this. We updated our CX5 to 16.26.1040 without issue. Fine. But we have two questions: First, the Release Notes were good, but are their any pointers to a user manual of the mlnx_* programs available under Ubuntu that we would actually run to configure the CX5 for Host Chaining? We couldn’t find that. So we’re totally lost as to where to start to set it up. … Second: A more detail question about Host Chaining that wasn’t 100% clear from the Release Note description. In our case, we have Layer 2 PTP Broadcast frames coming in on one port, and we want them hair-pinned out the other port. For example, a PTP Grandmaster’s L2 E2E (Announce,Sync,Follow-Up) all arrive on port A, and then leave on port B. At the same time, a PTP Slave’s (Delay_Request) should arrive on B and leave on port A. It feels like this is what Host Chaining can do; but the behavior with Ethernet L2 Multicast packets was not clear. Thanks again in advance. -Shep