Visit the post for more.
Overview
Facebook's new front-end server design focuses on maximizing compute capacity while minimizing power consumption. The redesign includes a collaboration with Intel to create a more efficient one-processor server, which significantly improves performance per watt compared to traditional two-processor servers.
What You'll Learn
1
How to optimize server design for better performance per watt
2
Why single-processor servers can outperform dual-processor setups in specific workloads
3
When to consider redesigning server infrastructure to accommodate new processor technologies
Prerequisites & Requirements
- Understanding of server architecture and processor design
- Experience with data center operations and server management(optional)
Key Questions Answered
How did Facebook redesign its web servers to improve performance?
Facebook redesigned its web servers to utilize a one-processor system with lower-power Intel Xeon-D CPUs, allowing for more CPUs per rack while maintaining the same power budget. This change improved performance per watt significantly compared to the previous two-processor systems.
What are the benefits of using Xeon-D processors in Facebook's data centers?
Xeon-D processors provide significant improvements in performance per watt, allowing Facebook to double the number of CPUs in each rack without exceeding the existing power budget. This approach helps avoid the flattening performance trajectory seen in previous generations of servers.
What challenges did Facebook face with traditional two-socket servers?
Traditional two-socket servers created a NUMA problem due to the QPI link between processors, which increased power consumption and limited scalability. As performance demands grew, the power limits of Intel CPUs became unsustainable, necessitating a redesign.
What is the role of Mono Lake and Yosemite in Facebook's server architecture?
Mono Lake serves as the server embodiment of the Xeon-D processor, designed for modularity and simplicity. Yosemite incorporates multiple Mono Lake boards and a multi-host NIC, allowing for efficient communication and bandwidth aggregation among CPUs.
Key Statistics & Figures
Power budget per rack
11kW
This power budget is crucial for determining the number of CPUs that can be accommodated in each rack.
CPU thermal design power (TDP)
65W
The new Xeon-D processors have a TDP of 65W, allowing for more CPUs per rack compared to the previous 120W processors.
Number of CPUs per rack
120 CPUs
With the new design, Facebook can fit 120 CPUs in a rack, doubling the previous capacity.
Technologies & Tools
Processor
Intel Xeon-d
Used in Facebook's redesigned server architecture to improve performance per watt.
Runtime
Hhvm
An open-source virtual machine used for executing programs written in Hack and PHP on Facebook's web servers.
Key Actionable Insights
1Consider transitioning to a one-processor server design to enhance performance per watt in your data center.This approach can lead to significant improvements in efficiency, especially in workloads that are heavily compute-bound, as demonstrated by Facebook's experience with Xeon-D processors.
2Collaborate with hardware partners to tailor processor designs that meet specific workload requirements.Facebook's partnership with Intel to develop the Xeon-D processor illustrates the benefits of aligning hardware capabilities with software demands, which can lead to innovative solutions and performance gains.
3Evaluate the memory bandwidth and capacity needs of your applications to avoid bottlenecks.Understanding the memory requirements of your workloads can help in selecting the right server architecture and processor, as Facebook found that their workloads were memory-bandwidth-bound rather than memory-capacity-bound.
Common Pitfalls
1
Relying on traditional two-socket server designs can lead to power inefficiencies and performance limitations.
As seen in Facebook's experience, pushing the power limits of dual-processor systems became unsustainable, highlighting the need for innovative designs that align with workload requirements.
Related Concepts
Server Architecture Optimization
Processor Design Collaboration
Data Center Power Efficiency