AMD server CPU 'EPYC' boosts Netflix server bandwidth to 380 Gbps

by Fritzchens Fritz

Netflix is trying to find the ideal hardware and software combination to improve the quality of video streaming. Such Netflix reports the result of using AMD 's server CPU ' EPYC'.

Serving Netflix Video at 400Gb / s on FreeBSD
(PDF file) https://people.freebsd.org/~gallatin/talks/euro2021.pdf

Netflix liefert 400 GBit / s pro Server aus: AMD EPYC ist Ampere und Intel überlegen --Hardwareluxx
https://www.hardwareluxx.de/index.php/news/hardware/prozessoren/57128-netflix-liefert-400-gbit-s-pro-server-aus-amd-epyc-ist-ampere-und-intel- ueberlegen.html

AMD's EPYC CPUs Push Netflix Server Bandwidth To 400 Gbps | Tom's Hardware
https://www.tomshardware.com/news/amd-epyc-cpus-netflix-bandwidth-400-gbps-per-server

Since 2020, Netflix has been delivering video content at a rate of 200 Gbps per server. Netflix's next goal is to double the speed to 400 Gbps, and to achieve this goal, we adopted AMD's second-generation server CPU 'EPYC 7502P'.

The EPYC 7502P has 32 cores, a maximum operating frequency of 2.5GHz, and a maximum operating frequency (when boosted) of 3.35GHz. Netflix's server is also equipped with 256GB DDR4-3200, and it is said that it will realize a bandwidth of about 150 Gbps with an 8-channel memory interface. In addition, 18 Western Digital SN720 2TB are used for each server as storage. We also introduced two Mellanox ConnectX-6 Dx network cards per server that connect via PCIe 4.0 x16. This configuration was expected to achieve speeds of 250 Gbps.

However, when I tested the connections on these servers, the first results showed a number of 240 Gbps. Seeing that memory bandwidth limits were below expectations, Netflix tried different NUMA (Non-Uniform Memory Access) configurations to maximize bandwidth. The NUMA configuration allows memory controllers and devices connected via PCIe to be assigned to the optimal core, but some processes between NUMA nodes can cause delays.

So Netflix can achieve speeds of up to 280Gbps by increasing the number of NUMA nodes to four and concentrating both storage and network-specific processes on NUMA nodes that are optimally connected to storage and PCIe devices. I did.

Software optimization was considered as the next step in reaching 400 Gbps. Implemented TLS on FreeBSD to avoid hardware bottlenecks. By offloading TLS encryption to the Mellanox ConnectX-6 Dx instead of the EPYC 7502P and adjusting it, we hit 190 Gbps for each Mellanox ConnectX-6 Dx adapter and up to 380 Gbps for the two. Processor utilization was 60% without NUMA and 50% with 4 NUMA nodes.

Also, even if Netflix changed the configuration to Intel 's Xeon Platinum 8352V or Ampere 's Altra Q80-30 and tested it, it was EPYC 7502P that showed the fastest value. Netflix has already set its next goal of reaching 800 Gbps, implying that it will announce new information in 2022.