R810 performance problem with 10GBit Interfaces

I have a throughput problem with 10GBit Ethernet interface cards in a Dell R810.

When running "iperf3" I get 5Gbit/s instead of the expected 9.4Gbit/s.
The same cards give 9.4GBit/s in a Dell R710.

System:

• Dell R810
• 4 CPUs Xeon E74820
• 251GB Ram
• 2x Dual 10GBit Intel 82599ES SFP+ in slots 3+4
  (moving them to slots 5+6 does not help)
• CentOS 7.4
• two 10GBit interfaces configured as bond0 (lacp)
• the servers are connected through a Mellanox switch

Commands:

• on R810 (server):
  

  iperf3 -s

• on R710 (client):
  

  iperf3 -c SERVER_IP

If I reverse the direction, I get the expected throughput of 9.4GBit/s

What I tried:

• increase MTU to 9000
• move 10GBit cards to slots 5+6
• pin iperf to a cpu
• pin interrupts to a cpu through
  "/proc/irq/N/smp_affinity"
  This increases the throughput from 4.8GBit/s to 5.2GBit/s
• increase ring parameters from 512 to 4096
  "ethtool -G p6p1 rx 4096 tx 4096"

The R810 are in a lab, and I can take them down at any time.

Any ideas?

Regards