Capacity Engineering

What is Performance? What is Capacity?

Latency. Throughput. Efficiency (cost). Scalability.

“The highest throughput, with acceptable latency, in the smallest footprint” – Performance and Capacity Engineering @ Facebook

Things to consider:

  • Architecture
  • Code base
  • Infrastructure
  • Hardware

Architecture

Most applications have several components. Those components perform different tasks and have different resource requirements. Some are data intensive, some of computational intensive, ect.

Horizontally Integration

Dedicated to communication between other subsystems.

Vertically Integration

The process of integrating subsystems according to their function. Quick, involves only the necessary vendors.

Top-Down Design

An overview of system is formulated without going into details for any part. Each part of it then refined into more details.

Bottom-Up Design

Individual parts of the system are specified in details. The parts are then linked to form larger components.

CPU- Central Processing Unit

More versatile than GPU’s, but not as fast. Individual CPU cores are faster and smarter than individual GPU cores, but not as a hold. Suited for a wide variety of workloads, especially with per-core performance.

Constructed from millions of transistors, the CPU can have multiple cores.

GPU - Graphical Processing Unit

GPUs can process data several order of magnitude faster than CPU due to parallelism. GPUs are better for repetitive and highly-parallel computing tasks. Excel in machine learning, simulations, and scientific computation.

A processor that is made up by many smaller and more specialized cores. By working together, the cores deliver massive performance because it is divided up and processed across many cores.

Analytics in Capacity engineering

Brendan Greegg’s Flame Graphs

“Without data, and with analysis of data, you’re not working on performance. You’re working on something else” – Bill Jia, Facebook

Machine learning needs a lot of computing. Sometimes your servers have fixed computation and memory.

Bill Jia (Facebook) seems to be the expert on this in the world.