Lecture 2-3: Cloud

Massive Scale
- If need to process large amounts of data or data processed by server is growing exponentially (ex: Facebook 30K in 2009 → 60K in 2010 → 180K in 2012) = USE CLOUD!
On-demand Access
- Think about it as renting a car vs. buying a car (you need to take care of the car)
- HaaS: rent barebones hardware machines, do whatever you want with them
- IaaS: compute & storage (ex: AWS EC2), made possible by virtualization and containerization (ex: Kubernetes, Docker)
- PaaS: compute & storage with software platform (ex: Google AppEngine host Python, Go)
- SaaS: essentially just applications (ex: Google Docs)
- FaaS: event-driven, short-lived, serverless, provided environment (ex: AWS Lambda)
Data-intensive Nature
- Computationally intensive programs like MPI-based or HPC need more compute power → these rely on supercomputers (more CPU power)
- Data intensive programs on the other hand don’t need much CPU power, the limitation is disk space and network because of processing large amounts of data → rely on cloud
Need for New Cloud Computing Paradigms
- Easy to run highly parallel programs because data can be sharded across servers that run the same operation
  - Ex: Google uses MapReduce, Yahoo and Facebook use Hadoop, etc.

Public (accessible by any paying customer) vs. Private (accessible only to company employees)
Data centers have nodes (servers) stacked on top of each other in racks, connected to each other via through their ToR (top of the rack) switches
- Servers
- Cooling
- Power [memorize these equations]
  - $WUE = \frac{\text{Annual Water Usage}}{\text{IT Energy}} L/kWh$
  - $PUE = \frac{\text{Facility Power}}{\text{IT Power}}$ → lower bound is 1.0 companies try to get as close to this as possible (ex: Google has 1.1)

Docker	Container	VM	Kubernetes
Uses containerization to build, test, deploy software	OS-level separation (share host OS, but not processes and resources)	Hardware-level separation using hypervisor, allows multiple OS’s and applications on same machine	Orchestration software used to automate deploying, scaling, managing clusters
An implementation of containers, but standardized to use Daemon environment, comes with ecosystem of applications like Docker Desktop	Can be programmed to use a customizable environment for special needs	Ex: different VMs in 425 each have their own OS and own local clock, etc.	Pods run code, nodes run multiple pods, cluster is a group of multiple nodes