> On Jan 20, 2020, at 4:47 AM, Hrvoje Popovski <[email protected]> wrote: > > On 16.1.2020. 18:45, Scott Cheloha wrote: >> Here's a first batch of conversions: rx refill timeouts for bnxt(4), >> myx(4), and vr(4). All of these can run during a softclock(). Will >> changing these to one tick break these drivers? > > Hi all, > > i tried this diff with myx and performance are the same as with plain > kernel. > > myx0 at pci4 dev 0 function 0 "Myricom Z8E" rev 0x01: msi, model > 10G-PCIE2-8BL2-2S, address 00:60:dd:45:ba:f8 > myx1 at pci5 dev 0 function 0 "Myricom Z8E" rev 0x01: msi, model > 10G-PCIE2-8BL2-2S, address 00:60:dd:45:ba:f9
Appreciate the testing. Given what dlg@ has said in the past I think there should only be a performance change in a livelock situation. What metric did you use to compare performance? I think you could induce a livelock with tcpbench(8) or iperf3 from ports. CC'd bluhm@, who would know more. To be clear, I'm of the opinion that we should not be spinning the softclock() unless the alternative is palpably reduced performance for a significant group of users. However, I don't want to tread on network performance without analysis and buy-in from the people who work on that, hence all the CCs. Probably missing a few CCs, but I figured I'd start with the API and driver authors.
