System specs for the Packet machines used in the Kubecon Seattle comparison

cncf / cnf-testbed

ARCHIVED: 🧪🛏️Cloud-native Network Function (CNF) Testbed --> See LFN Cloud Native Telecom Initiative https://wiki.lfnetworking.org/pages/viewpage.action?pageId=113213592

https://wiki.lfnetworking.org/pages/viewpage.action?pageId=113213592

Apache License 2.0

163 stars 51 forks source link

System specs for the Packet machines used in the Kubecon Seattle comparison #117

Closed taylor closed 5 years ago

taylor commented 5 years ago

This provides details on the machines used in the K8s and OpenStack clusters as well as those used for generating traffic.

[x] Add high-level specs for k8s worker nodes and openstack compute nodes
[x] Add high-level specs for traffic generator machine
[x] Add BIOS information for the machines running dataplane NFs
[x] Add detailed machine specs for worker/compute nodes

taylor commented 5 years ago

Worker/compute machines hosting CNF and VNFs

The Kubernetes and Openstack clusters will have 2-3 machines which run the network functions.

The system hardware configuration is based on the Packet m2.xlarge.x86.

The default dual port Mellanox ConnectX-4 NIC has been replaced by quad port Intel x710 NIC.. The NIC ports are connected to 10GbE ports on the top-of-rack switches.

Specs at a glance:

CPU: Dual socket Xeon Gold 5120 (2.2Ghz)
Memory: 384 GB of DDR4 ECC
Storage: 3.2 TB of NVMe Flash
NIC: Quad port Intel x710

Master/control machines

TBD

Traffic generator machine

Same as worker node - eg. m2.xlarge.x86 based machine with a quad port Intel x710 NIC

taylor commented 5 years ago

Additional NIC information https://github.com/cncf/cnfs/issues/94

taylor commented 5 years ago

@michaelspedersen, @pmikus, @mackonstan this ticket will contain the specs for the systems being used at Packet for the Kubecon comparison test case.

ahothan commented 5 years ago

Xeon Gold 5120 = 14 cores total of 28 cores (2 sockets) How are you planning to map these cores?

ahothan commented 5 years ago

keep 1/socket for host / kernel that leaves you 2x13 cores to play with, between VPP and NFs

michaelspedersen commented 5 years ago

keep 1/socket for host / kernel that leaves you 2x13 cores to play with, between VPP and NFs

This has also been my go-to approach for configuration so far. We've been trying out a few different configurations (scaling both VPP and VNFs/CNFs)

ahothan commented 5 years ago

in general it would be reasonable to find the "best performing system" given the same amount of resources and preferably in conditions that are realistic and close to production conditions.

taylor commented 5 years ago

BIOS details are in #129