Have been investigating less than expected performance when enabling SQM (Which I will raise as separate issue) but as part of the investigations it seems like the Brume 3/GL-MT5000 may not have optimal CPU affinity set for optimal performance when not using HW offload?
Looks like although the affinity is set to use all CPU cores it just sticks all interrupts on CPU0? I need to do some more experimentation to see if manual tweaks can better balance this out but in the meantime was curious what the intend behaviour is here. It does look like other platforms (MT3000/MT6000) have optimisations built into the irq-smp-affinity service?
Yes I see the same as you. The affinity is set for all cores but as your first screenshot shows they are all actually pegged to cpu0 and nothing is counting on cpu1-3 and all the stats are 0.
It feels like this is potentially not optimal? Perhaps there is a reason for this and performance is better in this existing configuration? Behaviour didn’t seem to match the all cores setting so was trying to understand the intent here. Thanks
The R&D team has confirmed that this behavior is normal, and the default interrupt binding already provides good overall performance.
That said, if anyone has alternative configurations that can deliver better performance across a wide range of scenarios, please feel free to share your findings and feedback. We’ll be happy to discuss them with the R&D team and evaluate whether improvements can be incorporated in future versions.
