Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

  • Alec Hothan I would prefer to see a bare metal server with dedicated NIC connected to the TOR(s) in order to be able to run SW traffic generator on this server. This cannot be achieved using VMs. The NIC should be in preference a NIC we are familiar to use and is known to work - e.g. Intel X710 family (XXV710 2x25G) - those are the NIC that are most used for SW traffic generators and we prefer to avoid troubleshooting untested HW profiles. We need 2 physical interfaces to the TOR(s). Please consider this as it is the necessary condition to run traffic gen tools like NFVbench on that RI. If you want to get fully automated testing on data plane we will also need to program the TOR in order to allow traffic to flow from the traffic gen to the compute nodes. We can discuss about the TOR side config offline with the installer team (this will depend also on hw the deployer configures the openstack dataplane)
    • Parker Berberian What kinds of CPU and RAM would be needed to run a traffic generator in software? I assume the specs listed below for the compute / controller machines would be ample.
    • Alec Hothan yes same as compute node would do
    • Xavier Grall imho, we should separate jump host and traffic generator function: jump host can run as VM, while SW traffic generator must run on a dedicated server (or "half" server, ie a socket/processor with associated RAM and NIC in a single numa zone) ; if Intel XXV710 are used, at least 2 NICs are required  
    • Alec Hothan jump host can be separate and live in a VM, but if you already have a bare metal server for data plane testing, why would you want to run jump host in a VM? For traffic generation, we need 2 ports interfaces, at least same speed as compute nodes. This can be done by adding a dedicated XVV710 NIC for data plane testing. 
      • Xavier Grall I am not familiar of OPNFV POD configuration and I only say that the jump host **can** run on a VM (if it is the usual setup) ; for XXV710, yes it is recommended to use the two 25GE ports on different NICs for testing a compute node having also two 25GE ports, even when they are on the same dual-port NIC 
        • Alec Hothan it really does not matter on the traffic gen side if the 2 ports are on same NIC or different NICs. From a practical point of view it is simpler to have a dedicated NIC and both ports on same NIC because that allows you to make that extra NIC optional (only required to run traffic generation) without impacting the remaining functions of that server. From an automation pov, we still need:
          •  a way to discover from the lab infra service which NIC/ports to use for traffic generation at the time the traffic generator code is installed on that server
          • how to configure the TOR side of those interfaces given that there is no guarantee they would be configured properly and the configuration of these interfaces will depend on the parameters of the openstack infra installed
        • Xavier Grall For the different XXV710 NICs requirement, you can look at this Intel document: https://fast.dpdk.org/doc/perf/DPDK_19_05_Intel_NIC_performance_report.pdfpdf#page=16. We have the same result in our lab in Orange when applied to TRex generator with a direct cable between 2 ports of a same XXV710: the line rate 25Gbps is never reached (even not with 1500B frames size).
          The opinion of an Intel contributor would be interesting: Trevor Cooper ?
          Regarding automation, which is a quite relevant question, beyond the port selection issue, there are several options for the generator installation:
          • Script-based on a specific machine, eg on the same machine as the jump host but with hardened performance-oriented configuration (isolcpu, huge pages…) so we should consider “jump host can be on the generator machine” (rather than “the generator can be on the jump host machine”).
          • Installer-based, by considering the generator as a specific NFVI role/function to deploy.
          • NFVI-based, eg using ironic if available, or in a VM or Container with dedicated ports (pci-passthrough or sriov PF or VF) if the generator can support it.

...