Results for the Poster submission to SIGCOMM 2025.
The tests were run on a Turris Omnia 2020, Microtik RouterBoard hEX RB750Gr3 and in Mininet. All test were executed on this topology:
Results from the test on the Microtik RouterBoard, TurrisOmnia, and in Mininet.
The results are averaged over 10 test runs.
The time where the error is introduced into the network is marked with the horizontal dottet line.
The MTU of the network is set to 850.
On the Turris Omnia, the drop in throughput is more severe and lasts longer in the nft case compared to the eBPF case.
This indicates that the device takes longer to resolve the routing error with nft.
Results from the test on the Microtik RouterBoard, TurrisOmnia, and in Mininet.
The results are averaged over 10 test runs.
The time where the error is introduced into the network is marked with the horizontal dottet line.
A noticeable spike in RTT occurs when the error is introduced,
with the delay being more pronounced in the nft test case on the Turris Omnia compared to the eBPF test case.
This suggests that the Turris Omnia requires more time to adjust routing rules using nft than with eBPF.
Results from the test on the Microtik RouterBoard, TurrisOmnia, and in Mininet.
The results are averaged over 10 test runs.
The time where the error is introduced into the network is marked with the horizontal dottet line.
The system load for the Mininet hosts is not included,
as it could not be measured accurately for the ShortCut implementation due to significant noise from various other processes running on the host.
All systems are packet forwarding near full load, with the Routerboard at a load of 2 and the Turris Omnia at a load of 4.
The different load values reflect the fact that the Turris Omnia has twice the number of CPU cores and can therefore can handle double the load.
A Load higher than the CPU counts shows that task are waiting before they can be executed. A load less than the CPU count indicates idle CPU time.
When an error is introduced, both the Turris Omnia and the Routerboard experience a temporary increase in load due to the additional computation required for rerouting.
In the nft test case, the Turris Omnia shows a higher and longer-lasting load spike, indicating that the nft implementation is less efficient than the eBPF implementation.