On 29.12.2015. 22:23, Hrvoje Popovski wrote: > On 29.12.2015. 17:49, Mark Kettenis wrote: >>> Date: Tue, 22 Dec 2015 23:45:49 +0100 >>>> >>>> On 22.12.2015. 22:08, Mark Kettenis wrote: >>>>>> Anybody willing to give this a spin? I don't have access to hardware >>>>>> currently... >>>>>> >>>>>> Thanks, >>>>>> >>>>>> Mark >>>> >>>> Hi, >>>> >>>> i'm sending 1.1Mpps and this patch almost immediately trigger OACTIVE >>>> flag. Patch is applied on clean source few minutes ago. If there is >>>> anything i can do more, please tell.... >> ok, that diff wasn't quite right. Here is a better one. This one >> gets rid of the ridiculous number of scatter segments on the 82598. >> There really is no point in allowing that many segments, and with the >> new code it would reduce the usable part of the tx ring significantly. >> >> I did some testing of forwarding performance on a machine with two >> sockets filled with: >> >> cpu0: Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz, 2400.37 MHz >> >> for a total of 16 cores forwarding between ix1 and ix0: >> >> ix0 at pci3 dev 0 function 0 "Intel X540T" rev 0x01: msi, address >> 0c:c4:7a:4d:a3:e4 >> ix1 at pci3 dev 0 function 1 "Intel X540T" rev 0x01: msi, address >> 0c:c4:7a:4d:a3:e5 >> >> I basically tested how many pps I could push through the box without >> loss, and many pps got through if sent 1Mpps into the box. All >> testing was done with pf disabled. >> >> With -current I got the following numbers: >> >> - 730kpps without loss >> - 82kpps when receiving 1Mpps >> >> and if I set net.inet.ip.ifq.maxlen to 8000 I got: >> >> - 740kpps without loss >> - 640-740kpps when receiving 1Mpps (fluctuating) >> >> With this diff I got: >> >> - 670kpps without loss >> - 250kpps when receiving 1Mpps >> >> and if I set net.inet.ip.ifq.maxlen to 8000 I got: >> >> - 690kpps without loss >> - 680kpps when receiving 1Mpps (fluctuating) >> >> So the maximal throughput goes slightly down, but it seems it with the >> diff it behaves betterunder load. >> >> Further tests are welcome! > > > Hi, > > i'm getting similar results with this patch. Sending 1Mpps and getting > 750Kpps. I didn't see OACTIVE flag even when generating 14Mpps :) > > PF=no > ddb.console=1 > kern.pool_debug=0 > net.inet.ip.forwarding=1 > net.inet.ip.ifq.maxlen=8192 > kern.maxclusters=32768 > > One cpu... > Intel(R) Xeon(R) CPU E5-2430 v2 @ 2.50GHz, 2800.01 MHz > > ix0 at pci2 dev 0 function 0 "Intel 82599" rev 0x01: msi, address > 90:e2:ba:33:af:ec > ix1 at pci2 dev 0 function 1 "Intel 82599" rev 0x01: msi, address > 90:e2:ba:33:af:ed >
just for fun i leave 14Mpps through box over night and ssh is alive, very slow but it's alive ... and i'm getting stable 650Kpps without OACTIVE flag ...
