On Thu, Jan 21, 2021 at 6:34 PM lejeczek via CentOS <centos at centos.org> wrote: > Hi guys. > > Hoping some net experts my stumble upon this message, I have > an IPoIB direct host to host connection and: > > -> $ ethtool ib1 > Settings for ib1: > Supported ports: [ ] > Supported link modes: Not reported > Supported pause frame use: No > Supports auto-negotiation: No > Supported FEC modes: Not reported > Advertised link modes: Not reported > Advertised pause frame use: No > Advertised auto-negotiation: No > Advertised FEC modes: Not reported > Speed: 40000Mb/s > Duplex: Full > Auto-negotiation: on > Port: Other > PHYAD: 255 > Transceiver: internal > Link detected: yes > > and that's both ends, both hosts, yet: > > > $ iperf3 -c 10.5.5.97 > Connecting to host 10.5.5.97, port 5201 > [ 5] local 10.5.5.49 port 56874 connected to 10.5.5.97 port > 5201 > [ ID] Interval Transfer Bitrate Retr Cwnd > [ 5] 0.00-1.00 sec 1.36 GBytes 11.6 Gbits/sec 0 > 2.50 MBytes > [ 5] 1.00-2.00 sec 1.87 GBytes 16.0 Gbits/sec 0 > 2.50 MBytes > [ 5] 2.00-3.00 sec 1.84 GBytes 15.8 Gbits/sec 0 > 2.50 MBytes > [ 5] 3.00-4.00 sec 1.83 GBytes 15.7 Gbits/sec 0 > 2.50 MBytes > [ 5] 4.00-5.00 sec 1.61 GBytes 13.9 Gbits/sec 0 > 2.50 MBytes > [ 5] 5.00-6.00 sec 1.60 GBytes 13.8 Gbits/sec 0 > 2.50 MBytes > [ 5] 6.00-7.00 sec 1.56 GBytes 13.4 Gbits/sec 0 > 2.50 MBytes > [ 5] 7.00-8.00 sec 1.52 GBytes 13.1 Gbits/sec 0 > 2.50 MBytes > [ 5] 8.00-9.00 sec 1.52 GBytes 13.1 Gbits/sec 0 > 2.50 MBytes > [ 5] 9.00-10.00 sec 1.52 GBytes 13.1 Gbits/sec 0 > 2.50 MBytes > - - - - - - - - - - - - - - - - - - - - - - - - - > [ ID] Interval Transfer Bitrate Retr > [ 5] 0.00-10.00 sec 16.2 GBytes 13.9 Gbits/sec > 0 sender > [ 5] 0.00-10.00 sec 16.2 GBytes 13.9 > Gbits/sec receiver > > It's rather an oldish platform which hosts the link, PCIe is > only 2.0 but with link of x8 that should be able to carry > more than ~13Gbits/sec. > Infiniband is Mellanox's ConnectX-3. > > Any thoughts on how to track the bottleneck or any thoughts Care to capture (a few seconds) of the *sender* side .pcap? Often TCP receive window is too small or packet loss is to blame or round-trip-time. All of these would be evident in the packet capture. If you do multiple streams with the `-P 8` flag does that increase the throughput? Google says these endpoints are 1.5ms apart: (2.5 megabytes) / (13 Gbps) = 1.53846154 milliseconds >