Hi Billy, Thank you for the reply. I can understand the experiment condition clearly.
>[[BO'M]] This is a valid suggestion and is worth investigating. I would like >to use linux tracing tools such as perf to see when or if the pmd thread in >testpmd is de-scheculed and if so for how long. Actually I’d start by checking >this on the testpmd thread on the host - as this removes any doubt about >timestamp accuracy on VMs. Very good! Actually, I have another suggestion to control CPU allocation for testpmd using taskset. In this time, we can observe the latency of process scheduling, and compare their pattern with the outlier. If the latency has a similar pattern, it would be a major cause of the outlier. See you in the meeting. Regards, Chunghan Lee ---------------------------------------------------------- Chunghan Lee Cloud Network Project. Fujitsu Laboratories Ltd. 4-1-1 Kamikodanaka, Nakahara-ku, Kawasaki-shi, Kanagawa 211-8588, Japan E-mail： lee.chung...@jp.fujitsu.com<mailto:lee.chung...@jp.fujitsu.com> ---------------------------------------------------------- From: O Mahony, Billy [mailto:billy.o.mah...@intel.com] Sent: Monday, January 30, 2017 7:12 PM To: Lee, Chunghan/李 忠翰; Gray, Mark D; therb...@redhat.com<mailto:therb...@redhat.com> Cc: opnfv-tech-discuss@lists.opnfv.org<mailto:opnfv-tech-discuss@lists.opnfv.org> Subject: RE: Discussion points (outlier) for PriPath Hi Chunghan, Some responses below. We can also discuss at the OVS-NFV synch-up call later today. /Billy. From: Lee, Chunghan [mailto:lee.chung...@jp.fujitsu.com] Sent: Monday, January 30, 2017 2:20 AM To: O Mahony, Billy <billy.o.mah...@intel.com<mailto:billy.o.mah...@intel.com>>; Gray, Mark D <mark.d.g...@intel.com<mailto:mark.d.g...@intel.com>>; therb...@redhat.com<mailto:therb...@redhat.com> Cc: Lee, Chunghan <lee.chung...@jp.fujitsu.com<mailto:lee.chung...@jp.fujitsu.com>>; opnfv-tech-discuss@lists.opnfv.org<mailto:opnfv-tech-discuss@lists.opnfv.org> Subject: Discussion points (outlier) for PriPath Hi, I considered that the reason why the outlier (such as, 120-200 or 200-400) for Hi Pri PTP is observed with 120% tput. There are two estimation for the outlier. The first one is VM scheduling latency. In this experiment, the VM is used for the measurement. The results would be different due to the number of VMs on node. If there are many VMs on the node and it suffers a lack of CPU resource, the scheduling latency will be increased and it would be a major cause of the outlier. Did you confirm the number of VMs in the experiment ? [[BO'M]] There is just one VM which is pinned to two host cores. However I have not isolated the those cpus so linux could decide to schedule something on them - however with a lightly loaded system and 28 other cores to choose from I don’t think this should be an issue. The next one is the packet processing delay of testpmd. In my understanding, testpmd is user process on VM. Although pmd thread forwards packets to testpmd as soon as possible, the context switching of testpmd would be occurred by kernel. As a result, testpmd cannot process the packets from the pmd thread as soon as possible. [[BO'M]] This is a valid suggestion and is worth investigating. I would like to use linux tracing tools such as perf to see when or if the pmd thread in testpmd is de-scheculed and if so for how long. Actually I’d start by checking this on the testpmd thread on the host - as this removes any doubt about timestamp accuracy on VMs. Please let me confirm the experiment condition without baseline again. In the priority path, is there no queue for Hi Pri PTP? Thus, there is only one designated queue for Lo Pri2. The traffic of Lo Pri2 is en/dequeued on the designated queue. There is no special function on DPDK OVS switch to process the traffic Hi Pri PTP. Is it right ? [[BO'M]] For the baseline case yes there is just one queue on all of the ports. (In reality the PMD polls queues not ports.) For the priority path case there are two queues on both the NIC ports. And the code sets up a flow-director filter on the NICs to assign the priority traffic to one of the queues. Regards, Chunghan Lee
_______________________________________________ opnfv-tech-discuss mailing list opnfv-tech-discuss@lists.opnfv.org https://lists.opnfv.org/mailman/listinfo/opnfv-tech-discuss
