Age | Commit message (Collapse) | Author | Files | Lines |
|
Some PODs (e.g. ericsson-virtual*) use more than 5000 x 2M hugepages,
together with 3G+ per-socket dpdk memory. Adjust our FDIO scenario
definitions to accomodate such configurations without triggering the
OOM.
Change-Id: Ibce2316f158bde98ad8e54f3eec75a827982d417
Signed-off-by: Alexandru Avadanii <Alexandru.Avadanii@enea.com>
|
|
Hugepage count has been recently bumped for virtual PODs via IDF
changes in Pharos, so align our FDio scenarios with the new RAM
requirements.
While at it, fix wrong pod_config template evaluation by moving it
after the templated scenario files are expanded, since pod_config
relies on scenario node definition.
Also, configure VPP to use decimal interface names by default to
align with Pharos macro for the VPP interface name string.
Change-Id: Ib3a89c294a3a2755567fdbe07e3be2b8ca1a5714
Signed-off-by: Alexandru Avadanii <Alexandru.Avadanii@enea.com>
|
|
NOTE: only os-nosdn-nofeature-noha is parameterized for now.
- move config drive & disk creation from prepare_vms to create_vms;
- make default disk size(s) configurable based on scenario defaults
and vPDF;
* compute nodes require 2 disks to be defined in vPDF, since the
pillar reclass model assumes /dev/vdb is reserved for cinder;
* if multiple disks are defined in vPDF, they are created and
attached accordinly (only ctl01 and cmp nodes are parameterized
in this change; only for the os-nosdn-nofeature-noha scenario);
- vCPU specifications are deduced based on vPDF (sockets, cores);
* threads/core is hard set to 2 since vPDF does not have a key
for it;
* NUMA resources are distributed evenly based on the number of
sockets configured in PDF;
* no less than the mininum requirement for a scenario is allocated
(e.g. if PDF specifies 2 cores, but the scenario requires at
least 4 cores, the larger value will be used);
- RAM is deduced based on PDF (but no less than the mininum req is
allocated, e.g. if PDF specifies 2GB RAM for computes, but the
scenario requires at least 8GB, the larger value will be used);
Change-Id: I97188aa2a1006865b8429eb6483e10c76795f7d2
Signed-off-by: Alexandru Avadanii <Alexandru.Avadanii@enea.com>
|
|
- cmp, gtw: bump RAM allocation to accomodate hugepages/VPP;
for now we overcommit, gtw01 resources can probably be lowered;
- submodule: add salt-formula-neutron so we can locally patch it;
- repo:
* FD.IO repos for VPP packages;
* networking-vpp PPA for python-networking-vpp Neutron driver;
- use vpp-router for L3, disable neutron-l3-agent;
- baremetal_init: apply repo config before network (otherwise UCA
repo is missing when trying to install DPDK on baremetal nodes);
- arm64: iommu.passthrough=1 is required on ThunderX for VPP on
newer kernels;
Design quirks:
- vpp service runs as 'neutron' user, which does not exist at the
time VPP is installed and initially started, hence the need to
restart it before starting the vpp-agent service;
- gtw01 node has DPDK, yet to configure it via IDF we use the
compute-specific OVS-targeted parameters like
`compute_ovs_dpdk_socket_mem`, which is a bit misleading;
- vpp-agent requires ml2_conf.ini on ALL compute AND network nodes
to parse per-node physnet-to-real interface names;
- vpp process is bound to core '1' (not parameterized via IDF);
Change-Id: I659f7dbebcab7b154e7b1fb829cd7159b4372ec8
Signed-off-by: Alexandru Avadanii <Alexandru.Avadanii@enea.com>
|
|
While at it, rename FDIO (VPP) scenarios to align with OPNFV FDS
and OPNFV Apex projects.
Change-Id: I9aab5dc4a0dc41a2cc996687a8a2726d03288678
Signed-off-by: Alexandru Avadanii <Alexandru.Avadanii@enea.com>
|