Back of the envelope Calculations

Servers	RAM	Processor	Storage
Web	Low	High	Low
Application	High	Medium	Medium
Storage	Low	Medium	High

Facebook open-sourced its data center design in 2011

Web Servers - 32GB RAM, 500GB Storage, For high end computations - Custom 16 core processors (partnered with Intel)

Application Servers - RAM up to 256GB, and Storage (Disk + Flash) - up to 6.5TB

Storage Servers, youtube example

• Blob Storage - for its encoded videos

• A temporary processing queue storage that can hold a few hundred hours of video content uploaded daily to YouTube for processing

• Specialized storage called Bigtable for storing a large number of thumbnails of videos.

• RDBMS - Relational database management system for users and videos metadata

Typical Server Specifications

No. of sockets - 2

Processors - Intel Xeon X2686

No. of cores - 36 cores (72 threads)

RAM - 256 GB

Cache L3 - 45 MB

Storage Capacity - 15TB

Important Latencies

Component	Time (nanoseconds)
L1 cache reference	0.9
L2 cache reference	2.8
L3 cache reference	12.9
Main memory reference	100
Compress 1KB with Snzip	3,000 (3 microseconds)
Read 1MB sequentially from memory	9,000 (9 microseconds)
Read 1MB sequentially from SSD	200,000 (200 microseconds)
Roundtrip within the same data center	500,000 (500 microseconds)
Read 1MB sequentially from SSD with speed ~1DB/sec SSD	1,000,000 (1ms)
Disk seek	4ms
Read 1MB sequentially from disk	2ms
Send packet SF->NYC	71ms

Important Rates / QPS

QPS = Queries per second

QPS Handled by	QPS
MySQL	1,000
Key-value store	10,000
Cache server	100,000 - 1M

Requests estimations

RPS = Requests per second

• CPU bound requests

• Memory bound requests

RPS_CPU = NUM_CPU / Task time

RPS Memory = RAM_SIZE / (WORKER_MEMORY X TASK_TIME)

Eg. for RPS_CPU, Assuming each tasks takes 200ms and the CPU has 72 threads

RPS_CPU = 72 / 200ms = 360RPPS

Eg. for RPS Memory, Assuming Each worker needs 300MB, and takes 50ms for a request for a server with 240GB RAM,

240GB/(300MB * 50MS) = 16000RPS