Rack-Level Cost and Power Savings

Server racks in data centers deliver features and scale related to critical applications such as network virtualization, security, load balancing and telemetry. The number or VMs that can be deployed per server in the rack relates to available CPU resources and networking bandwidth. The Netronome® Agilio® solution used in compute node servers improves both metrics, and significantly reduce CAPEX and power consumption per rack, while boosting output performance and scalability per rack.

Benchmarks

A common performance target for VMs in servers is approximately 1Mpps per VM. Using a 16-core server, an Agilio-accelerated server will allow 15 VMs to be deployed on the system, each VM using one core. The remaining core is reserved for overhead. With this configuration and a rack capacity of 20 servers, the expected output is 300Mpps for the entire rack, which costs roughly $55K in CAPEX. A total of five racks would be required to achieve similar output with non-accelerated servers, resulting in a CAPEX of $250K. The bottleneck that causes this cost disparity lies in the performance of the vSwitch. Relying on CPUs for vSwitching tasks limits performance and consumes excessive cores. The Agilio solution eliminates the dependency, resulting in significant reductions in rack-level TCO.

Test Setup/Tools

The key inputs for calculating TCO for a rack include server and power costs, ToR switch power and costs, and expected output per VM. Assumptions for modeling TCO are vSwitch performance for non-accelerated and Agilio-accelerated OVS. The assumptions in the calculator are derived from rigorous testing done in labs (see Netronome’s OVS test plan). The calculator reveals savings based on the CAPEX and OPEX of an accelerated and non-accelerated rack configuration that drives on an expected application output, measured in Packets Per Second (PPS).

Architecture

Many modern data centers adopt a leaf/spine switching topology, with much of the complex networking tasks such as tunneling and quality of service being pushed to the server. This analysis focuses at the rack-level for this topology. A ToR leaf switch provides 40G connectivity to 20 servers within the rack. Each server uses Open vSwitch to provide VXLAN tunneling and security policies for multiple VMs in each server.