>-----Original Message----- >From: Bokhan Artem [mailto:a...@eml.ru] >Sent: Friday, October 17, 2014 4:53 AM >To: Fujinaka, Todd; E1000-devel@lists.sourceforge.net >Subject: Re: [E1000-devel] performance of two 82599 nics at >a host > >Thank you for the answer, Todd. > >15.10.2014 23:07, Fujinaka, Todd пишет: >> Can you send us the output of "lspci -vvv" (run as root) >and ethtool -s? >Attached. > >> I'm guessing that you're trying to use memory on remote >NUMA nodes. You can try running: >> https://software.intel.com/en-us/articles/intel- >performance-counter-monitor-a-better-way-to-measure-cpu- >utilization >You are right. When using both adapters remote memory is >used. Remote memory is not used when any of two adapters run alone. >Results are in attachment too. Why this happen?
This list strips attachments, so I can't see your results, but in general you can see that the driver allocates as many queues as CPUs. Because of this you will have queues that are on both NUMA nodes. The idea behind the default behavior of the driver is to use the CPU which is local to the user space application, not necessarily the CPU that is local to the NUMA node of the device. Performance degradation can be far greater when feeding application from a remote queue even if the queue is local to the interface. You can find multiple threads regarding NUMA performance if you look at the archives of this mailing list. For optimal performance we generally recommend to map the affinity of the queues to CPUs using the set_irq_affinity script provided with the driver. You will have to disable irqbalance. Aside from your results I don't know anything about the actual test, or the type of traffic or your setup, so it's hard to tell where your bottleneck is or if it’s related to NUMA at all. In general when reporting issues I recommend to open a bug on sourceforge as it makes it easier to upload files and for us to track it. Thanks, Emil ------------------------------------------------------------------------------ Comprehensive Server Monitoring with Site24x7. Monitor 10 servers for $9/Month. Get alerted through email, SMS, voice calls or mobile push notifications. Take corrective actions from your mobile device. http://p.sf.net/sfu/Zoho _______________________________________________ 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
