Thank you for answer. You are quite right. I have LVS machine and TCP traffic. So it's more like router. And while testing with iperf traffic distributed among 24 RX queues. Now I understand why it's happens.
Few more questions: 1. With Six-Core CPUs and HT I get 24 TxRx queues and only 16 of them are loaded. Are there any advantages of using Hyper-Threading technology or better turn it off and get 12 loaded queues? 2. Do you plan to implement more than TxRx 16 queues for RSS? Or maybe there are some other technologies useful for multicore routers? 3. Is it possible to equally distribute traffic to all 24 queues using Perfect Filters? Considering that clients can come from various ip addresses. Aleksey -----Original Message----- From: Jesse Brandeburg [mailto:[email protected]] Sent: Monday, November 21, 2011 7:35 PM To: Aleksey Chudov Cc: [email protected] Subject: Re: [E1000-devel] ixgbe rx_queue On Sun, 20 Nov 2011 15:22:03 -0800 Aleksey Chudov <[email protected]> wrote: > I have two Six-Core CPUs and ixgbe driver creates 24 tx/rx queues. But > traffic distributed only among 16 RX queues. > > How to direct traffic to all 24 rx queues? If you are using only UDP traffic, (you didn't mention the traffic) then you will be directed to a queue only via the RSS mechanism, which only can use 16 queues. It appears you are likely routing? flow director really only works if you're using the server as a TCP endpoint. if you see lots of fdir_miss in the counters you're falling back to queue picking with RSS. We have some UDP flow director code that is not usable for production because it doesn't work with ip fragments. The other option you have is to use perfect filters of some kind (and a newer kernel) via ethtool ntuple if you have a way to direct certain ranges of addresses or ports to certain queues (depends on your workload). Jesse ------------------------------------------------------------------------------ All the data continuously generated in your IT infrastructure contains a definitive record of customers, application performance, security threats, fraudulent activity, and more. Splunk takes this data and makes sense of it. IT sense. And common sense. http://p.sf.net/sfu/splunk-novd2d _______________________________________________ E1000-devel mailing list [email protected] https://lists.sourceforge.net/lists/listinfo/e1000-devel To learn more about Intel® Ethernet, visit http://communities.intel.com/community/wired
