Mellanox 40Gbp and 100Gb NIC cards issue RoS v7.11

Hi guys

I am having some troubles in testing some routerOS x86_64 hardwares.

we have setup 2 dell servers R620 and R420..

setup R420 2 E5-2470v2 cpus.. for full capacity + mellanox MXC354-FCBT dual port 40GB (with firmwware upgradede and switched to ethernet mode)
setup2 R620 with 2 E5-2697v2 cpus + Mellanox MCX455-ECAT single port 100GB (with firmware upgraded latest available on nvidia site, and also in ethernet mode.

So i have tested with DAC 40GB cable.. with autonegotiation off.. it will not link 40GBps full-duplex.. so i removed the dac and placed 40gb gbic with cable and it linked full duplex 40gbps on both servers..

so i did some testing with bandwith test server.. and TCP running both sides.. i can only get 15Gbps aggregate traffic.. cpu wise running average 20% to 30% on both machines.. it seems to be something on the network side so i tried out.. jumbo frame set 9800 on both ethernet adapters as shown on pictures below.. but i cannot get past 15gbps..

if i do internal core test local on the server i can get nearly over 60gbps internal on the cpus TCP packets.. but on the network interfaces cannot get past 15gbps aggregate.. wondering if its software wise or hardware wise on the servers?

also multi x86_64 is enabled on both servers..


Any ideas on what else i could try out?
Both servers are PCIe 3.0 slot cards.. so i am wondering if its hardware limitation.

Or could it be a limitation on the 24h trial license? as we are using trial 24h license to test the hardware.
Server test1.png
server test2.png
server test3.jpeg

Ok Guys

so i did some more testing and it looks like X16 is 64GBps max… on 8 x 8 gbps.. just need to figure out on how to fine tune this based on the CPUs cores 48threads and 2nd server 40threds it should work fine.. so far i could get around 34Gbps aggregate speeds on TCP using 100G NIC on one side but linked up in 40Gbps e and using dual port 40gbps on the 2nd server..
UPGRADE2.jpeg
UPGRADE3.jpeg
UPGRADE1.jpeg

Ok so more tests done..

Memory upgrade on servers to 78GB R420 and 96GB on R620…

somehow it increased the speeds as expected we are now getting 64Gbps aggregate troughput between servers
server 64gbps.png

More RAM likely means more channels in use to the socket and so more memory bandwidth.