A vector size less than 2 indicates a ton of headroom. You shouldn’t drop packets - except due to I/O bottlenecks - at a vector size significantly less than 150.
Rough, off-the cuff guesses [for production images, on “reasonable” hardware]: vector size < 2 should be good for O(100kpps). Vector size 150 should be good for O(5mpps). Thanks… Dave P.S. The clocks/pkt figures below look like they may be from a TAG=vpp_debug image... Or not. It’s hard to tell at 20 KPPS... From: SAKTHIVEL ANAND S [mailto:[email protected]] Sent: Tuesday, July 25, 2017 9:20 AM To: Dave Barach (dbarach) <[email protected]> Cc: chenxndsc <[email protected]>; vpp-dev <[email protected]> Subject: Re: [vpp-dev] 回复: vpp cpu usage utility Thanks Dave.. I am on vpp16.06 and after adding my nodes I get the following in “vppctl show r”. One thing I observe (as expected) is when I push more traffic to see higher "Vectors/Call" the Clocks value of nodes reduces. Now question is neither clocks or Vectors/Call indicate that still some processing power is left. By the way, what is prevailing vector size and per-node vector size in the below output? < last 128 main loops 0.00 per node 0.00> What are these values and why these are zeros? Thanks in advance. root@ubuntu:~# vppctl show run Time 1678.2, average vectors/node 1.45, last 128 main loops 0.00 per node 0.00 vector rates in 1.9800e4, out 1.9337e4, drop 4.6368e2, punt 6.9123e-2 Name State Calls Vectors Suspends Clocks Vectors/Call GigabitEthernet13/0/0-output active 117 117 0 2.71e3 1.00 GigabitEthernet13/0/0-tx active 117 117 0 1.94e4 1.00 GigabitEthernet1b/0/0-output active 22450924 32449638 0 1.26e3 1.45 GigabitEthernet1b/0/0-tx active 22450924 32449638 0 3.64e3 1.45 arp-input active 21064 22591 0 2.61e4 1.07 dpdk-input polling 388349175 33227874 0 2.54e4 .09 dpdk-process any wait 0 0 560 2.09e5 0.00 error-drop active 709544 778129 0 2.25e3 1.09 error-punt active 116 116 0 1.16e4 1.00 ethernet-input active 27238 28772 0 5.35e3 1.06 interface-output active 88 88 0 4.86e3 1.00 ip4-arp active 15 16 0 3.63e4 1.07 ip4-classify active 1600 1603 0 1.04e4 1.00 ip4-input-no-checksum active 22854016 33199102 0 1.79e3 1.45 ip4-local active 22841621 33173678 0 1.29e3 1.45 ip4-lookup active 45295061 65626189 0 1.51e3 1.45 ip4-lookup-multicast active 10723 22451 0 1.95e3 2.09 ip4-miss active 12074 23802 0 1.37e3 1.97 ip4-rewrite-transit active 22450827 32449541 0 1.47e3 1.45 ip4-udp-lookup active 22841621 33173678 0 1.29e3 1.45 llc-input active 3393 3394 0 4.01e3 1.00 mynode_lookup active 22841621 33173678 0 1.89e3 1.45 mynode_input active 22841621 33173678 0 1.49e3 1.45 unix-epoll-input polling 365372525 0 0 3.36e3 0.00 -------------------------- -Sakthivel S OM On Mon, Jul 17, 2017 at 11:25 PM, Dave Barach (dbarach) <[email protected]<mailto:[email protected]>> wrote: The “show runtime” command displays the prevailing vector size, and per-node vector sizes. Those stats are reasonably equivalent to a vector engine load average. Thanks… Dave From: [email protected]<mailto:[email protected]> [mailto:[email protected]<mailto:[email protected]>] On Behalf Of SAKTHIVEL ANAND S Sent: Monday, July 17, 2017 12:57 PM To: chenxndsc <[email protected]<mailto:[email protected]>>; vpp-dev <[email protected]<mailto:[email protected]>> Subject: Re: [vpp-dev] 回复: vpp cpu usage utility Thanks chenxndsc for the response. Yes, the polling shall take the 100% CPU. My question is related to know how much more load a core can take more. For e.g. Lets say along with normal dpdk polling, if a core processeses few sessions, how much CPU it takes in processing sessions apart from polling. This is required to know how many sessions a can handle and hence how many cores are needed to handle desired number of sessions? -Sakthivel Ss On 17-Jul-2017 21:08, "chenxndsc" <[email protected]<mailto:[email protected]>> wrote: hello ANAND 100% is the real cpu utilization, because VPP use POLL mode, if you know how DPDK works, you gotta to know it better. 来自我的华为手机 -------- 原始邮件 -------- 主题：[vpp-dev] vpp cpu usage utility 发件人：SAKTHIVEL ANAND S 收件人：vpp-dev 抄送： Hi I am working on an use case where i just want to measure CPU usage of vpp task. Linux "TOP" command shows 100% cpu always (even though there is no active processing in it,since it is running on closed/tight loop). Is there any way to measure "real/actual" cpu usage of VPP task? My setup: Ubuntu 16.04 with VPP16.06 -- Thanks in advance Sakthivel S OM -- Thanks Sakthivel S OM
_______________________________________________ vpp-dev mailing list [email protected] https://lists.fd.io/mailman/listinfo/vpp-dev
