BackendDevOpsOther

OCP vs. Traditional Server Systems

Last Updated on: 17th June 2024, 05:04 pm

Introduction

RTB House is utilizing two types of infrastructure in our Data Centers: Traditional and Open Compute Project. The Open Compute Project (OCP) is an initiative to design and openly share hardware designs for data centers and other computing infrastructure. The project aims to increase the efficiency, flexibility, and scalability of data center hardware while reducing its environmental impact and cost. One of the design improvements in OCP is a new approach to power delivery. Instead of each server being equipped with two PSUs (power supply units), all servers in the rack are powered by power supplies on the rack level, using a redundant power bus bar.

Hardware specification

  1. OCP Gigabyte TO25-Z11-AA01-000
    CPU: 1x EPYC-9684X
    RAM:1.5TB DDR5
    NET: BCM57508-N2100G (NetXtreme-E Dual-port 100G QSFP56 Ethernet OCP 3.0 Adapter)
  2. ASUS RS520A-E12-RS24U
    CPU: 1x EPYC-9684X
    RAM:1.5TB DDR5
    NET: BCM57508-P2100G (NetXtreme-E Dual-port 100G QSFP56 Ethernet PCIe4.0 x16 Adapter)
  3. ASUS RS520A-E12-RS12U
    CPU: 1x EPYC-9684X
    RAM:1.5TB DDR5
    NET: BCM57508-P2100G (NetXtreme-E Dual-port 100G QSFP56 Ethernet PCIe4.0 x16 Adapter)

Server #1 is an OCP system powered by a busbar (OCP provides power through the busbar to all servers in the rack). Server #2 and Server #3 are equipped with standard PSUs (two per system).

Load and performance

All three systems have the same CPU model. On our platform, we are balancing load on specific nodes based on CPU. Nodes with the same CPU have the same load, regardless of server brand and design. All three servers are stressed the same way within 24 hours (CPU frequency rises when it is not under heavy load):

The moving average of CPU frequency on all cores.

Temperatures

Server #1 (yellow) has an ~15 degrees lower CPU temperature than Server #2 and Server #3 (green):

CPU temperatures

Power usage

Server #1 (yellow) power use is 50-100W lower than Server #2 and Server #3.

Power consumption

Power tweaks

Out of the box, ASUS servers had even higher energy use. We have adjusted power curves for them, sacrificing lower temperatures for lower FAN speed, hence lower power used for FANs. This way, approximately 70W of energy was saved.

Power draw of ASUS servers before the tweaks

Power draw of ASUS servers after the tweaks

Conclusions

OCP architecture provides better energy efficiency than traditional design without compromising on performance. Less energy used means lower temperatures. Both of those can have a significant influence on the operating cost of a data center. In our case, this is one of the most important benefits of the OCP platform, among many others:

  • Computing power density.
  • Modular design: it is very easy to remove and install hardware components, including hot-swapping storage, which reduces time and money spent on maintenance
  • Server management tools are based on trusted and tested solutions.

Comments are closed.

More in Backend