To conduct these tests i booted friendlywrt from an sd card.
I disabled the firewall, removed the bridge and brought up the 2.5Gbe ports as dhcp clients.
I used friendlywrt 21.02 for test purposes as I had already seen similar real world metrics when testing with Debian and the newer version of friendlwrt.
My test source/network is a recent thinkpad x1 with a realtek 8156 2.5Gbe usb3 adapter connected through a qnap 2.5gbe switch.
As the results below show testing these ports with `iperf3` suggests a total 2Gbp/s+ non-duplex max throughput.
Testing with actual data copying across the network by ssh however suggests far lower performance, and in fact in some cases, worse performance than the onboard 1Gbp/s port.
SSH tests were done with a single block 1GB disk image dd'ed from `/dev/zero` held in `tmpfs`, and sshed to `/dev/null`
Test results
eth0 1Gbe
this device appears to support full duplex operation altho other tests may be impacted by the limitations of the source interface (ie realtek adapter)
iperf3 900Mbits/s+ both directions
iperf3 bidir = 1.8Gbits/s+ total
scp down avg 45.1MB/s
scp up avg 15.6MB/s
eth1 2.5Gbe
iperf3 2Gb/s+ both directions
iperf3 bidir = 2Gb/s+ total
scp down avg 41.5MB/s
scp up avg 23.3MB/s
eth2 2.5Gbe
iperf3 2Gb/s+ both directions
iperf3 bidir = 2Gb/s+ total
scp down avg 41.1MB/s
scp up avg 23.8MB/s
Comparison to odroid n2+ + realtek 8156 adapter
iperf3 2Gb/s+ both directions
iperf3 bidir = 2Gb/s+ total
scp down avg 121.3MB/s
scp up avg 128.5MB/s
Conclusion
At least in one direction the 1Gbp/s port is consistently outperforming the 2.5Gbp/s ports
These tests were far from scientific and dont test throughput when more than one of the ports are in use at the same time. I suspect (and my anecdotal experience is) that this would make whatever performance issues these devices have worse.
I have had similar experiences in (very preliminary) testing of an odroid m1 which has the same chipset - im wondering if the issue is down to the capacity of the controllers and whether 2.5gbe is really possible on these boards in any meaningful way
I disabled the firewall, removed the bridge and brought up the 2.5Gbe ports as dhcp clients.
I used friendlywrt 21.02 for test purposes as I had already seen similar real world metrics when testing with Debian and the newer version of friendlwrt.
My test source/network is a recent thinkpad x1 with a realtek 8156 2.5Gbe usb3 adapter connected through a qnap 2.5gbe switch.
As the results below show testing these ports with `iperf3` suggests a total 2Gbp/s+ non-duplex max throughput.
Testing with actual data copying across the network by ssh however suggests far lower performance, and in fact in some cases, worse performance than the onboard 1Gbp/s port.
SSH tests were done with a single block 1GB disk image dd'ed from `/dev/zero` held in `tmpfs`, and sshed to `/dev/null`
Test results
eth0 1Gbe
this device appears to support full duplex operation altho other tests may be impacted by the limitations of the source interface (ie realtek adapter)
iperf3 900Mbits/s+ both directions
iperf3 bidir = 1.8Gbits/s+ total
scp down avg 45.1MB/s
scp up avg 15.6MB/s
eth1 2.5Gbe
iperf3 2Gb/s+ both directions
iperf3 bidir = 2Gb/s+ total
scp down avg 41.5MB/s
scp up avg 23.3MB/s
eth2 2.5Gbe
iperf3 2Gb/s+ both directions
iperf3 bidir = 2Gb/s+ total
scp down avg 41.1MB/s
scp up avg 23.8MB/s
Comparison to odroid n2+ + realtek 8156 adapter
iperf3 2Gb/s+ both directions
iperf3 bidir = 2Gb/s+ total
scp down avg 121.3MB/s
scp up avg 128.5MB/s
Conclusion
At least in one direction the 1Gbp/s port is consistently outperforming the 2.5Gbp/s ports
These tests were far from scientific and dont test throughput when more than one of the ports are in use at the same time. I suspect (and my anecdotal experience is) that this would make whatever performance issues these devices have worse.
I have had similar experiences in (very preliminary) testing of an odroid m1 which has the same chipset - im wondering if the issue is down to the capacity of the controllers and whether 2.5gbe is really possible on these boards in any meaningful way