Thanks for your reply. I'm not having any issues with 82579 and I'm not having RPS on that host. A bit outdated e1000 is in use with MSI-X and RSS support. Everything was working straight away. I even do not know how data sheet looks like.
The fact that RSS is Microsoft`s spec, just proving again, that Microsoft is very weird company. However, I do not understand, why to use RSS from Microsoft instead of updating it to something more relevant. I'm not the single person having issues with 82599, BTW. Also, it's like a good idea to update documentation on this a bit. Haven't found anywhere in documents, anything about "if you have PPPoE, PPTP, L2TP... use RPS to balance traffic. Our hashing functions useless here". Conclusion could be made from reading whole document completely, understanding how RSS works in 82599, but that's a lot of time. In any case will see how RPS will do RSS work. Hopefully overhead is not big. Thank you all for suggestions. ----- Original Message ----- > From: "Todd Fujinaka" <todd.fujin...@intel.com> > To: "Jack Spinov" <spi...@timegroup.ae> > Cc: e1000-devel@lists.sourceforge.net > Sent: Friday, May 16, 2014 9:18:49 PM > Subject: RE: [E1000-devel] ixgbe: how to balance PPPoE traffic via RSS > to multiple queues > > So the offsets are off because you are tunneling and it looks like > you're coming from the same port and address. > > To answer some of your other questions: > > RSS is defined by Microsoft and you'll have to take that up with > them. > > You've now been given two options to balance your traffic: vmdq (not > trivial) and RPS. > > If you're not having trouble with the 82579 you may already be using > something to spread the traffic as e1000e only has one queue and so > does the 82579. > > Todd Fujinaka > Software Application Engineer > Networking Division (ND) > Intel Corporation > todd.fujin...@intel.com > (503) 712-4565 > > -----Original Message----- > From: Jack Spinov [mailto:spi...@timegroup.ae] > Sent: Friday, May 16, 2014 12:22 AM > To: Skidmore, Donald C > Cc: e1000-devel@lists.sourceforge.net > Subject: Re: [E1000-devel] ixgbe: how to balance PPPoE traffic via > RSS to multiple queues > > Sorry, givven offsets in different dimensions. > > PPPoE src IP starts at 184 bit, dst IP starts at 216, instead of 96 > and 128 respectively. > > Just to correct myself. > > ----- Original Message ----- > > From: "Jack Spinov" <spi...@timegroup.ae> > > To: "Donald C Skidmore" <donald.c.skidm...@intel.com> > > Cc: e1000-devel@lists.sourceforge.net > > Sent: Friday, May 16, 2014 11:14:36 AM > > Subject: Re: [E1000-devel] ixgbe: how to balance PPPoE traffic via > > RSS to multiple queues > > > > Thanks for your reply, Donald. > > > > Looks like you're right. Src IP offset starts at 23 bit and Dst IP > > at > > 27. Instead of 12 and 16. > > > > Just a rhetoric question: why not to have hash function on MAC? > > > > And not rhetoric: how to balance damn PPPoE traffic? > > > > BTW I didn't had any issue balancing such traffic on 82579. > > Everything worked perfectly. > > > > ----- Original Message ----- > > > From: "Donald C Skidmore" <donald.c.skidm...@intel.com> > > > To: "Jack Spinov" <spi...@timegroup.ae> > > > Cc: e1000-devel@lists.sourceforge.net > > > Sent: Friday, May 16, 2014 12:00:10 AM > > > Subject: RE: [E1000-devel] ixgbe: how to balance PPPoE traffic > > > via > > > RSS to multiple queues > > > > > > I might be misunderstanding what you're trying to do here, but > > > the > > > reason you're going to queue 0 is that you are tunneling over > > > PPPoE. > > > Because of this all of the offsets that the HW expects are no > > > longer correct so it can't hash on the IPv4, IPv4_TCP and if UDP > > > fields RSS needs. > > > > > > Thanks, > > > -Don Skidmore <donald.c.skidm...@intel.com> > > > > > > > -----Original Message----- > > > > From: Jack Spinov [mailto:spi...@timegroup.ae] > > > > Sent: Thursday, May 15, 2014 5:46 AM > > > > To: Skidmore, Donald C > > > > Cc: e1000-devel@lists.sourceforge.net > > > > Subject: Re: [E1000-devel] ixgbe: how to balance PPPoE traffic > > > > via > > > > RSS to multiple queues > > > > > > > > According to driver source code functions are executed in order > > > > ( > > > > I skipped > > > > IPv6 ): > > > > IPv4, IPv4_TCP and if UDP flow hashing is enabled - IPv4_UDP. > > > > > > > > If yes, then I do not understand, why hash function returns 0 ( > > > > i.e. assigns packets to 0 queue ) for PPPoE packets. Looks like > > > > a > > > > bug to me. > > > > > > > > ----- Original Message ----- > > > > > From: "Jack Spinov" <spi...@timegroup.ae> > > > > > To: "Donald C Skidmore" <donald.c.skidm...@intel.com> > > > > > Cc: e1000-devel@lists.sourceforge.net > > > > > Sent: Thursday, May 15, 2014 3:11:03 PM > > > > > Subject: Re: [E1000-devel] ixgbe: how to balance PPPoE > > > > > traffic > > > > > via RSS to > > > > multiple queues > > > > > > > > > > And one more thing. According to Data Sheet, RSS can properly > > > > > parse tunneled packets, treating them as IP packets. In this > > > > > case hash is built by SourceAddress and DestinationAddress. > > > > > However, it's stated, that in case RSS function failed to > > > > > parse > > > > > packet - function returns zero. Without ntuple and Flow > > > > > Director > > > > > ( do not work for tunneled packets ), this means, that > > > > > packets > > > > > will always go to queue 0, which is my case. > > > > > > > > > > So the question is: is there a way to find out what functions > > > > > RSS uses to parse IP packets, i.e. value of MRQC register? > > > > > And > > > > > modify it. > > > > > It > > > > > seems to me, that needed function for parsing is just not > > > > > included into check. Adding it will render RSS index assigned > > > > > and packets will be properly distributed among queues. > > > > > Unicorns > > > > > and rainbows. > > > > > > > > > > > > > > > > > > > > > > ----- Original Message ----- > > > > > > > From: "Jack Spinov" <spi...@timegroup.ae> > > > > > > > To: "Donald C Skidmore" <donald.c.skidm...@intel.com> > > > > > > > Cc: e1000-devel@lists.sourceforge.net > > > > > > > Sent: Thursday, May 15, 2014 12:37:47 PM > > > > > > > Subject: Re: [E1000-devel] ixgbe: how to balance PPPoE > > > > > > > traffic via > > > > > > > RSS to multiple queues > > > > > > > > > > > > > > Thanks for your reply. > > > > > > > > > > > > > > I'll try that for synthetic tests. But it's unclear what > > > > > > > to > > > > > > > do, in case I will not have VLANs or will have just few > > > > > > > in > > > > > > > production? > > > > > > > VMDq > > > > > > > without virtual environment looks like workaround to me. > > > > > > > > > > > > > > Let's say server is BRAS with PPPoE server terminating > > > > > > > client sessions. Without PPPoE - everything is working as > > > > > > > it > > > > > > > should. > > > > > > > Once > > > > > > > PPPoE in effect - everything gets to queue 0. > > > > > > > > > > > > > > How can I configure VFs in this case? > > > > > > > > > > > > > > ----- Original Message ----- > > > > > > > > From: "Donald C Skidmore" <donald.c.skidm...@intel.com> > > > > > > > > To: "Jack Spinov" <spi...@timegroup.ae>, > > > > > > > > e1000-devel@lists.sourceforge.net > > > > > > > > Sent: Wednesday, May 14, 2014 9:05:27 PM > > > > > > > > Subject: RE: [E1000-devel] ixgbe: how to balance PPPoE > > > > > > > > traffic via > > > > > > > > RSS to multiple queues > > > > > > > > > > > > > > > > Maybe try VMDq which could sort by L2 (MAC address and > > > > > > > > VLAN). > > > > > > > > > > > > > > > > Thanks, > > > > > > > > -Don Skidmore <donald.c.skidm...@intel.com> > > > > > > > > > > > > > > > > > -----Original Message----- > > > > > > > > > From: Jack Spinov [mailto:spi...@timegroup.ae] > > > > > > > > > Sent: Tuesday, May 13, 2014 7:39 AM > > > > > > > > > To: e1000-devel@lists.sourceforge.net > > > > > > > > > Subject: [E1000-devel] ixgbe: how to balance PPPoE > > > > > > > > > traffic via RSS to multiple queues > > > > > > > > > > > > > > > > > > Hello, everyone. > > > > > > > > > > > > > > > > > > Have spent a lot of time trying to balance traffic > > > > > > > > > into > > > > > > > > > multiple RSS queues, but traffic falls into same > > > > > > > > > queue > > > > > > > > > all the time. I've read all the documents, threads I > > > > > > > > > could find, but still cannot find a way to solve my > > > > > > > > > problem. > > > > > > > > > > > > > > > > > > My configuration: 3 servers with 82599 adapters. > > > > > > > > > Connected > > > > > > > > > like > > > > > > > > > this: > > > > > > > > > > > > > > > > > > Packet generator ( PG )<--- 10G ---> Packet router ( > > > > > > > > > PR > > > > > > > > > )<--- > > > > > > > > > 10G > > > > > > > > > ---> Packet > > > > > > > > > dumper ( PD ) > > > > > > > > > > > > > > > > > > PG connects to PR via PPPoE and generates packets > > > > > > > > > directed to Packed dumper ( PG ). > > > > > > > > > > > > > > > > > > PR running 3.2.23 kernel and latest ixgbe 3.21.2. PG > > > > > > > > > running > > > > > > > > > 3.12.18 and same > > > > > > > > > latest ixgbe 3.21.2, with default settings. I've > > > > > > > > > generated traffic using iperf, netperf, ping. For > > > > > > > > > iperf, > > > > > > > > > netperf, both TCP and UDP - all goes to default queue > > > > > > > > > - > > > > > > > > > 0, while the rest 15 are idle. As on receiving > > > > > > > > > interface, as on sending for PR. In case of using > > > > > > > > > ping - > > > > > > > > > everything is distributed properly. > > > > > > > > > Which is > > > > > > > > > another riddle for me. > > > > > > > > > > > > > > > > > > I've tried using macvlan ( different MACs on PR, PG > > > > > > > > > ), > > > > > > > > > vlans ( different source IPs for PPPoE sessions, > > > > > > > > > different destination IP for PD ) - nothing helps. I > > > > > > > > > cannot use Flow Director as packets in tunneling are > > > > > > > > > not > > > > > > > > > analyzed by it ( as per DS ), this is proved by > > > > > > > > > ethtool > > > > > > > > > counter for fdir* variable. > > > > > > > > > > > > > > > > > > So what are my options? And how to guarantee, that in > > > > > > > > > production environment I will not face the same > > > > > > > > > issue? > > > > > > > > > Or > > > > > > > > > probably I'm doing something wrong? > > > > > > > > > > > > > > > > > > Thanks in advance. > > > > > > > > > > > > > > > > > > -------------------------------------------------------- > > > > > > > > > ------ > > > > > > > > > ---------------- "Accelerate Dev Cycles with > > > > > > > > > Automated > > > > > > > > > Cross-Browser Testing - For FREE Instantly run your > > > > > > > > > Selenium tests across 300+ browser/OS combos. > > > > > > > > > Get unparalleled scalability from the best Selenium > > > > > > > > > testing platform available Simple to use. Nothing to > > > > > > > > > install. > > > > > > > > > Get > > > > > > > > > started now for free." > > > > > > > > > http://p.sf.net/sfu/SauceLabs > > > > > > > > > _______________________________________________ > > > > > > > > > E1000-devel mailing list > > > > > > > > > E1000-devel@lists.sourceforge.net > > > > > > > > > https://lists.sourceforge.net/lists/listinfo/e1000-devel > > > > > > > > > To learn more about Intel® Ethernet, visit > > > > > > > > > http://communities.intel.com/community/wired > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > > ------------------------------------------------------------------------------ > "Accelerate Dev Cycles with Automated Cross-Browser Testing - For > FREE Instantly run your Selenium tests across 300+ browser/OS > combos. > Get unparalleled scalability from the best Selenium testing platform > available Simple to use. Nothing to install. Get started now for > free." > http://p.sf.net/sfu/SauceLabs > _______________________________________________ > E1000-devel mailing list > E1000-devel@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/e1000-devel > To learn more about Intel® Ethernet, visit > http://communities.intel.com/community/wired > ------------------------------------------------------------------------------ "Accelerate Dev Cycles with Automated Cross-Browser Testing - For FREE Instantly run your Selenium tests across 300+ browser/OS combos. Get unparalleled scalability from the best Selenium testing platform available Simple to use. Nothing to install. Get started now for free." http://p.sf.net/sfu/SauceLabs _______________________________________________ E1000-devel mailing list E1000-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel® Ethernet, visit http://communities.intel.com/community/wired
