If the VM does not use a DPDK application, there are VM exits when forwarding packets, which kills performance. Also, segmentation/receive offloads (GRO, GSO) are not supported by the DPDK vrouter in 3.0 (supported from 4.0), so that would also reduce the TCP throughput.
Raja From: Users <users-boun...@lists.opencontrail.org> on behalf of kevin parrikar <kevin.parker...@gmail.com> Date: Wednesday, July 19, 2017 at 11:57 AM To: dev <dev@lists.opencontrail.org>, "<us...@lists.opencontrail.org>" <us...@lists.opencontrail.org> Subject: [Users] dpdk vrouter performance hello All, i have contrail 3.0.* with DPDK vrouter running on 2 compute nodes with 1Gbx100 hugepages and 6 dedicated CPU for the DPDK vrouter process in 1 NUMA node. Created 2xUbuntu (16.04) 40Gb hugepage enabled vm with 6 dedicated VCPU in same compute nodes (Jumbo frames enabled) when performing iperf between the vms i get a through put of 2 Gbps where as vms on Non DPDK nodes i am getting 9Gbps. I assume vms running on DPDK nodes are not optimized and probably i need to compile in DPDK Libraries and DPDK application to get improved performance on Ubuntu16.04. My question is : with out DPDK libraries in guest why is the performance too low.Since qemu process is set to share memory and i assume with this option all packets ll be copied to guest memory by the dpdk vrouter. Thanks for your help Regards, Kev
_______________________________________________ Dev mailing list Dev@lists.opencontrail.org http://lists.opencontrail.org/mailman/listinfo/dev_lists.opencontrail.org
