Improving support for DPDK on FABRIC with UEFI and huge pages

[Ilya Baldin]

Dear experimenters,

We’ve received a number of requests to improve the support for DPDK on FABRIC. We are currently testing several of them in our development environment. We wanted to describe the changes we are working on here and provide the opportunity for feedback from you before we perform this change on the production environment. The change we are testing is a change to the default behavior of PCI passthrough and we want to make sure your experiments aren’t adversely affected. Our current testing so far does not suggest any such effects.

The change is driven by DPDK need to use huge pages and also by some Mellanox driver requirements that require UEFI (vs BIOS) boot support.

In order to make this change to FABRIC a number of underlying systems will need to be reconfigured:

1. Currently FABRIC VMs use qemu emulation of an older i440fx PCI chipset and this would be changed to use q35 PCIe chipset for all VMs going forward across all sites

2. The default boot option for all VMs will be changed from BIOS to UEFI

3. Huge pages will be configured on server nodes with CX-5, CX-6 network cards as well as FPGAs so they can be passed through to the VMs on those nodes

4. New VM flavors supporting huge pages will be added

#3 and #4 are of particular interest to us.

For #3 DPDK examples indicate huge pages of 2M and 1G, and since we have to preconfigure these on the servers, are these the appropriate sizes for us to configure?

For #4, as we cannot make all possible VM flavors huge-page capable due to resource constraints, the most likely scenario is a small number of instance flavors which combine a particular number of cores with some RAM and local disk sizes and are also huge-page enabled. If you have suggestions for those instance sizes that would be most useful to you, we are open to hear them.

Please respond on this thread if you have questions or suggestions.

[Author]

Posts