> On Sep 26, 2017, at 8:09 AM, Damien Clabaut <damien.clab...@corp.ovh.com> > wrote: > > Hello Keith and thank you for your answer, > > The goal is indeed to generate as much traffic per machine as possible (we > use pktgen-dpdk to benchmark datacenter routers before putting them on > production). > > For this we use all available CPU power to send packets. > > Following your suggestion, I modified my command to: > > ./app/app/x86_64-native-linuxapp-gcc/pktgen -l 0-7 -- -m 1.0 -s > 0:pcap/8500Bpp.pcap
I just noticed you were sending 8500 byte frames and you have to modify Pktgen to increase the size of the mbufs in the mempool. I only configure the mbufs to 1518 byte buffers or really 2048 byte, but I only deal with 1518 max size. The size can be changed, but I am not next to a machine right now. > > The issue is still reproduced, though will slightly lower performance > (reaching linerate at 8500 Bpp does not require much processing power) > > #sh int et 29/1 | i rate > 5 seconds input rate 36.2 Gbps (90.5% with framing overhead), 0 packets/sec > 5 seconds output rate 56 bps (0.0% with framing overhead), 0 packets/sec > > Regards, > > PS: Sorry for answering you directly, sending this message a second time on ML > > > On 2017-09-25 09:46 PM, Wiles, Keith wrote: >>> On Sep 25, 2017, at 6:19 PM, Damien Clabaut <damien.clab...@corp.ovh.com> >>> wrote: >>> >>> Hello DPDK devs, >>> >>> I am sending this message here as I did not find a bugtracker on the >>> website. >>> >>> If this is the wrong place, I would kindly apologize and ask you to >>> redirect me to the proper place, >>> >>> Thank you. >>> >>> Description of the issue: >>> >>> I am using pktgen-dpdk to replay a pcap file containing exactly 1 packet. >>> >>> The packet in question is generated using this Scapy command: >>> >>> pkt=(Ether(src="ec:0d:9a:37:d1:ab",dst="7c:fe:90:31:0d:52")/Dot1Q(vlan=2)/IP(dst="192.168.0.254")/UDP(sport=1020,dport=1021)/Raw(RandBin(size=8500))) >>> >>> The pcap is then replayed in pktgen-dpdk: >>> >>> ./app/app/x86_64-native-linuxapp-gcc/pktgen -l 0-7 -- -m [1-7].0 -s >>> 0:pcap/8500Bpp.pcap >> This could be the issue as I can not setup your system (no cards). The >> pktgen command line is using 1-7 cores for TX/RX of packets. This means to >> pktgen to send the pcap from each core and this means the packet will be >> sent from each core. If you set the number of TX/RX cores to 1.0 then you >> should only see one. I assume you are using 1-7 cores to increase the bit >> rate closer to the performance of the card. >> >>> When I run this on a machine with Mellanox ConnectX-4 NIC (MCX4), the >>> switch towards which I generate traffic gets a strange behaviour >>> >>> #sh int et29/1 | i rate >>> 5 seconds input rate 39.4 Gbps (98.4% with framing overhead), 0 >>> packets/sec >>> >>> A capture of this traffic (I used a monitor session to redirect all to a >>> different port, connected to a machine on which I ran tcpdump) gives me >>> this: >>> >>> 19:04:50.210792 00:00:00:00:00:00 (oui Ethernet) > 00:00:00:00:00:00 (oui >>> Ethernet) Null Unnumbered, ef, Flags [Poll], length 1500 >>> 19:04:50.210795 00:00:00:00:00:00 (oui Ethernet) > 00:00:00:00:00:00 (oui >>> Ethernet) Null Unnumbered, ef, Flags [Poll], length 1500 >>> 19:04:50.210796 00:00:00:00:00:00 (oui Ethernet) > 00:00:00:00:00:00 (oui >>> Ethernet) Null Unnumbered, ef, Flags [Poll], length 1500 >>> 19:04:50.210797 00:00:00:00:00:00 (oui Ethernet) > 00:00:00:00:00:00 (oui >>> Ethernet) Null Unnumbered, ef, Flags [Poll], length 1500 >>> 19:04:50.210799 00:00:00:00:00:00 (oui Ethernet) > 00:00:00:00:00:00 (oui >>> Ethernet) Null Unnumbered, ef, Flags [Poll], length 1500 >>> >>> The issue cannot be reproduced if any of the following conditions is met: >>> >>> - Set the size in the Raw(RandBin()) to a value lower than 1500 >>> >>> - Send the packet from a Mellanox ConnectX-3 (MCX3) NIC (both machines are >>> indentical in terms of software). >>> >>> Is this a known problem ? >>> >>> I remain available for any question you may have. >>> >>> Regards, >>> >>> -- >>> Damien Clabaut >>> R&D vRouter >>> ovh.qc.ca >>> >> Regards, >> Keith >> > > -- > Damien Clabaut > R&D vRouter > ovh.qc.ca > Regards, Keith
