Server Architecture

1. Scalable Computing/Analysis

1.1 Three architecture ?

1.1.1 Shared-Nothing versus Shared-Everything

Patterns for Distributed or Parallel computing
big problems, break it down, computing unit
which pattern to use
trader-off between compute and communication costs

MPP?

1.1.2 Shared-Memory

single machine, multiple processes
Apply in:

GPU-based parallel-computing architecture
Deep Learning

multiple machines
RDMA (Remote direct memory access)
data-center network
warehouse-scale computing

NUMA?

1.1.3 Shared-Storage

in 80’s, NFS/DFS/Andrew
today, SAN/NAS/Virtualized storage

2. Core elements in architecture

架构核心要素，思维导图
架构核心五要素

Ref Book:
大型网站技术架构

3. Server Architecture

3.1 服务器三大体系结构

Three architecture for business server

Symmetric Multi-Processing (SMP) 主流服务器架构

对称多处理架构：多个CPU对称工作，无主次或从属关系

Share same physical memory, access to any address takes same time
Thus is also called: 一致存储访问结构
Uniform Memory Access (UMA)  
Feature: sharing
Extension: more memory, faster CPU, more CPU, more i/o, more device
Defect: extension limited
Best practice: 2-4 CPUs

Massively Parallel Processing (MPP)

海量并行处理结构: 多个SMP服务器（节点）互联而成
完全无共享结构(Share Nothing)
Thus：扩展能力最好，目前512个节点，数千个CPU

Non-Uniform Memory Access (NUMA)

非一致存储访问结构
Feature: 多个CPU模块，每个模块多个CPU
独立内存，I/O
通过互联模块（eg. crossbar switch) 连接和信息交互
支持上百个CPU
Defects：远地内存延时太多
系统性能无法线性增加

MPP vs. NUMA

Criteria	MPP	NUMA
节点互联机制	I/O	inside same physical server
内存访问机制	access only local	access to all, remote delays
Usage	data warehouse, i/o	OLTP, little data, fast processing

服务器三大体系SMP、NUMA、MPP介绍

Transitioning from SMP to MPP, the why and the how

Multi core basics: AMP and SMP

Cluster Hardware Architecture

3.2 多核系统三大架构

一个操作系统的实例可以同时管理所有CPU内核，且应用并不绑定某一个内核

Asymmetric Multi-Processing (AMP)

非对称多处理架构
每个CPU内核可能运行一个独立的操作系统或同一操作系统的独立实例（instantiation）

AMP vs. SMP

Criteria	AMP	SMP
耦合	松耦合多CPU系统	紧耦合多CPU系统
OS	one CPU, one OS	multi CPU, share one OS
Apply	less	more

Bound Multi-Processing (BMP)

混合多处理架构
一个操作系统的实例可以同时管理所有CPU内核，但每个应用被锁定于某个指定的核心

3.3 补充

Seemingly, SMP(UMA), NUMA and MPP refers to CPU and RAM architecture

And SMP, AMP, BMP refers to CPU and OS architecture

More on 17-09-29:
SMP: multiprocessor computer hardware and software architecture
UMA: a shared memory architecture used in parallel computers
UMA may use bus-based SMP arch or crossbar switches

Intel 用了很牛的技术 SMT （同步多线程），一个CPU同时可以执行两个线程，如双核CPU可以有4个逻辑处理器

SMT(Simultaneous Multi-threading)
Def: Permits multiple independent threads to execute simultaneously on the SAME core.
Weaving together multiple “threads” on the same core

SMT v.s. Hyper Threading
SMT: an architecture feature of some processors
allows multiple thread to issue instructions on each cycle(same time ?)
Hyper-Threading is a specific implementation of SMT.

4 Design of System

4.1 Building block

decomposition and combination
分解和组合

System bus
系统总线
separate producer and consumer
Routing
路由
external route: different input –> |dispatch| –> component
internal route: service level, eg.consul, way like pattern matching

Mind data
Identify its ID before reaching Bus, called data normalization/data encapsulation

Queue
A special version of bus + route[optional] + storage ?
amends for the speed collapse of consumer and producer
1. Pub/Sub
  An candidate for route
  Managed by Publisher
2. Protocol
  To communicate though bus or queue, we need a protocol to define behaviour.
  like an SLA, service level agreement
4.2 vital services
1. sweeping
  代谢
  exception close, restart, restore…
  eg. supervisor/child process in erlang, supervision tree(like nginx)
2. High Availability
3. Security
  divide into initiative and passive(主动安全与被动安全)
  Initiative: authentication/authorization + TLS + encrypt + minimum input/output
  Passive: firewall
Overdraft protection
透支保护
auto scaling/back pressure/Service degradation

Reference:
软件设计杂谈