On Tue, May 15, 2018 at 04:51:53PM -0400, Adam Wallis wrote: > On 5/15/2018 11:07 AM, Greg Kroah-Hartman wrote: > > On Tue, May 15, 2018 at 09:53:57AM -0400, Adam Wallis wrote: > > Does this really do anything? Given the speed of USB3 at the moment, > > does fixing the memory to the node the PCI device is on show any > > measurable speedups? Last I remember about NUMA systems, it wasn't > > always a win depending on where the irq came in from, right? > > > > thanks, > > > > greg k-h > > -- > > To unsubscribe from this list: send the line "unsubscribe linux-usb" in > > the body of a message to majord...@vger.kernel.org > > More majordomo info at http://vger.kernel.org/majordomo-info.html > > > > I was getting really inconsistent throughput speeds on a system I was testing > with NUMA nodes. Using an SMMU in identity mode, I was able to track down > where > the performance deltas were coming from...Some of the rings were going to the > "wrong" node. > > Yes, it's possible to handle your IRQs with CPUs on the wrong NUMA node...but > I > would argue that it's always best to have the rings for USB controller X as > close to controller X if possible. Users can then properly constrain IRQs, and > even kernel threads to the right Domain if they so desire. > > After setting the IRQ affinity to the right node AND applying this patch, I > started getting much more reliable (and faster) results.
Ok, fair enough, I was hoping that "modern" systems would have better NUMA memory interconnects. I guess that isn't the case still :( thanks, greg k-h -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
