https://rancher.com/ logo
Join Slack
Powered by
This message was deleted.
# harvester
a
This message was deleted.
a
10G cards with BCM57416 chipset, firmware is up2date, cards have pci4 x8 (which should be around 15G/s)
b
Could be getting throttled on the switch or links between switches.
I am not a network person, but I think you might need to mention where you're running your iperf tests from and what kind of switches and configs you're using.
It would also be good to note what version of Harvester this is as pre 1.7 is 
wicked
and it moved to 
networkManager
after that.
a
iperf from inside a VM to a VM on the other host
running harvester 1.7.1
Cisco C9300-24UX
this is an empty cluster in PoC state
a
What MTU settings are you using? Do you have jumboframe support on your switches, and set the packet MTU settings to something sufficiently high (e.g. 9000 bytes) when setting up your nodes? Fragmentation can be a performance killer for iPerf runs.
a
MTU currently unset therefore 1500
will test 9000/jumbo frames
b
Likewise the VMs might need to be on the underlying VLAN and not the built in management network (which I think defaults to 1495 MTU or something)
Being on the underlying VLAN + 9000 like the hosts are are probably what you'd need for better performance.
a
yes vm traffic and storage traffic are on eperate 10G bonds/channels
a
a
will check the stack up from switch to harvester bond to vlan to vm config
b
Cluster Network + Virtual Machine Network.
a
To verify if MTU is indeed a contributing factor, you could re-run your iPerf test with a lower MTU value.
4 Views
Open in Slack
PreviousNext
Powered by