Distributed Systems. Today. Next. Distributed systems The class Models and architectures. A brief, gentle intro to Go

Size: px

Start display at page:

Download "Distributed Systems. Today. Next. Distributed systems The class Models and architectures. A brief, gentle intro to Go"

Steven Owens
6 years ago
Views:

1 Distributed Systems Today Distributed systems The class Models and architectures Next A brief, gentle intro to Go Image from

2 Welcome! What is this all about? A course on distributed systems (DS) With an experimental systems approach Focused on large-scale issues Our goals Learn about DS principles & current work Learn Go a some-what new programming language Build some interesting DS Learn/practice reading research papers 2

3 What is a distributed system? Very broad definition Set of independent, interconnected processors that communicate and coordinate their action by exchanging messages appears to its users as a single coherent system: DNS, The Web, BitTorrent, FB Consequences of the definition Concurrent execution is the norm so coordination is important No global clock and limits to the accuracy you can get when synchronizing multiple clocks Independent failures all systems can fail, distributed systems fail in new ways 3

4 Motivations and everyday examples Why do you want one? Resource sharing physical resources and information Computation speedup Reliability Communication Many applications are naturally distributed Some clear examples Finance and commerce e-commerce, paypal, digital currency Information society web, web search, user-generated content Arts and entertainment On-line gaming, music and video on demand Transport and logistics smart roads, cars, maps Environmental management sensor-based monitoring of natural phenomena 4

5 Trends driving distributed systems Mobile and ubiquitous computing Price/performance/size ratios in computer technology Pervasive networking Large number of nodes Increasing penetration Increasing bandwidth Content sharing Distributed systems as a utility an old idea on a whole new dimension 5

Cheaper & faster machines IBM 704 Scientific

not sold but rented at $50,000/month Apple

or 10MB HDD Motorola 68000, 5Mhz IBM AT 80286,

5 MIPS on the 10MHz model $9,600 (2015 $)

6 Cheaper & faster machines IBM 704 Scientific Computing, k instructions per second, not sold but rented at $50,000/month Apple Lisa, 1983 $24,000 (2015 $) 1MB RAM Optional 5 or 10MB HDD Motorola 68000, 5Mhz IBM AT 80286, MB Main Memory 20MB HD 1.5 MIPS on the 10MHz model $9,600 (2015 $) Apple iphone 6, GB Storage A8 dual-core 64b and M8 chips $399 6

7 Driving DS An interconnected world 29 Oct 1969 ARPANET first two interconnected nodes 1982 less than 1000 hosts in ARPANET Partial map of the Internet, 2005 The Opte Project 4-node ARPANET, circa billion people uses the Internet And that s only ~34% of World s population 7

8 Information growth >12x Growth in six years on information created, captured and replicated Zettabytes (3,600 Exabytes) 1 Exabyte = bytes, 1,000,000 TB 2007 (281 Exabytes) Predicted to reach 20 Zettabytes by 2020 J. Gantz, Digital Universe Study, IDC, EMC,

9 Large-scale distributed systems 1.2 billion in Facebook 250 million active users in Twitter >50 million Skype users online 9

10 Distributed systems challenges Heterogeneity networks, computer hardware, OS, programming languages, implementations Openness following agreed upon rules for extensibility, portability, interoperability Making resources available/accessible Security comes with sharing Failures handling detecting, masking, recovering Scalability in numbers and geographic span Transparency hide that the system is distributed, at least most of the time 10

11 Challenges and false assumptions* The network is reliable You need redundancy The network is secure Thread modeling and security from the start! The network is homogenous No proprietary protocols The topology does not change Location transparency, to a point *Based on Peter Detusch (1994) and James Gosling (1997) 11

12 Challenges false assumptions* Latency is zero Caching and replication and the speed of light Bandwidth is infinite Bandwidth is like memory Transport cost is zero Marshaling costs, transport assumptions, There is one administrator Resource access, firewalls, dealing with failures *Based on Peter Detusch (1994) and James Gosling (1997) 12

13 Why taking this class? You will learn Core concepts of distributed systems Abstractions, algorithms, implementation issues About popular distributed systems used by large companies today Google s File system Logical clocks Sun RPC and NSF How to build a real distributed system yourself! Through 2-3 large projects 13

14 Course and topic in context EECS 340 Networking Network protocols, routing protocols, Distributed systems: make use of networks EECS 343 Operating systems Resource management for single systems Distributed systems: management of distributed resources EECS 345 Distributed systems Explore solutions to DS problems and challenges Infrastructure software to help build DS and applications 14

15 Overview of this course 1. Introduction to Go 2. Networking, communication and overlays 3. Distributed file systems 4. Naming 5. Content distribution networks 6. Time and synchronization 7. Coordination and global state 8. Consensus, replication and fault tolerance 9. Mobile distributed systems 15

16 A word about requirements We expect you to have taken EECS 213 Intro to computer systems and EECS 343 Operating system or 340 Intro to networking Some concepts you will need Processes, threads, synchronization, IPC Basics of network programming Some skills you will need Hacking on a Linux/Unix-like environment Solid working experience with C This is a course with a heavy project component! Don t take it if you lack the prerequisites First project should give you a hint 16

17 Some practical notes Textbooks and papers A reference book Some classic and some research papers Homework assignments as reader enforcers Projects in a second An open take-home final Yes, grading Homework assignments 20% Class participation/presentations 15% Project 40% Exam 25% 17

18 Projects Using Go, a language well suited for DS Three parts Project 0 1 week Should you continue? Project 1 & 2 6 weeks Implement the Kademlia DHT Project 3 3 weeks Build a storage system for selfdestructive objects on top of your DHT First project on your own The rest in teams of 2-3 students (3 is preferred, 1 is not allowed) 18

19 Exam A take-home, final given in the last week of class Take-home, open book Example questions With asynchronous RPC, a client is blocked until its request has been accepted by the server. To what extent do failures affect the semantic of asynchronous RPC? DNS is susceptible to denial of service attacks, one of the papers discussed in class addresses this problem; describe the problem, explain the key ideas behind the authors solutions and their evaluation approach 19

20 Back in 5 20

21 Models and architectures Organization as the key to master complexity Two ways to look at distributed systems organization Software architecture / architectural models how to organize the system s software components, how components interact System architecture / physical models how to instantiate the set of components on real machines Fundamental models provide another set of abstraction lenses to focus on particular aspects for Interaction Failures Security 21

22 Architectural elements Communicating entities Systems perspective processes, threads or nodes Programming perspective objects, components or web services Communicating paradigms Inter-process communication Remote invocation RPC, RMI Object Indirect communication pub/sub, distributed shared memory Method call Object Object Object Referential decoupling (no need to know your name) Event-based, publish-subscribe Component Component Event bus Component Component Component Shared data space Temporal decoupling (no need to be around) Shared data spaces 22

23 System architectures Roles and responsibilities / system architecture As processes (objects, components, ) interact, they take on given roles that define the architecture: client-sever or P2P Guided by performance and reliability requirement of each component A classification based on the symmetry of the interaction between components Client-server and its variations Peer-based architecture Hybrid models 23

24 Client-server Basic idea There are processes offering services There are other processes using them Client and servers can be on different machines Interaction follows a request/reply model The underlying connection can use a connectionless (UDP) or connection-oriented protocol (TCP) Client Wait for reply Request Reply Server Time 24

25 Layers and tiers Server for some and client of others A typical three-layer model for database access User-interface UI part of an application Processing functionality of an application without data Data data to be manipulated through the application Responsible for ensuring data is persistent and consistent Many times a relational database User interface User-interface level Keyword expression Page with list A simple example an Internet search engine Database query Query generator HTML generator Ranking algorithm Ranked list of pages Processing level Database with Web pages Web pages titles with meta-information Data level 25

26 Multi-tiered architectures Three logical layers with multiple instantiations Over two or more tiers How much to place on each side? Thinner or fatter clients Fat s good Short(er) response times and leverage clients resources Thin s good Easy to manage Thin clients Client machines User interface User interface User interface User interface User interface Application Application Application User interface Database Application Application Application Database Database Database Database Database Server machines 26

27 Peer-to-peer we are all (mostly) equal Horizontal rather than vertical distribution Unstructured P2P systems Peers connect with other random peers Semi-structured models (superpeers) for scalability Super peer Regular peer 27

28 Structured models E.g. Distributed Hash Tables Peers connection and content mapping in Chord [13,14,15] 15 0 Actual node 1 [0,1] Data item with key k is mapped to the node with smallest identifier id >= k (successor(k)) [8,9,10,11,12] [2,3,4] [5,6,7] Associated data keys 28

29 Hybrid models Edge-server systems typical of content distribution networks (CDNs) Clients get content through an edge server Peer-assisted CDNs ISP ISP Internet ISP Edge server Content provider 29

30 Fundamental models Interaction Focused on fundamental properties, making explicit all relevant assumptions, to reason about Interaction Synchronous - Time to execute step of a process, transmit a msg and the drift rate of clocks have known upper bounds Asynchronous no bounds But is asynchronous realistic? Is there any send that will take >1260 secs (longest delay to Mars)?! Not the point General solutions that won t fail if the assumptions (delay) fail According to Fisher, Lynch and Paterson, there s no algorithm for the consensus problem in an asynchronous system 30

Fundamental models Interaction Process interaction through message exchanges Communication performance is often a limiting characteristic No global clock or global notion of

31 Fundamental models Interaction Process interaction through message exchanges Communication performance is often a limiting characteristic No global clock or global notion of time Clocks drift at different rates Logical clocks maybe enough X 10 send Meeting? Re: Meeting? Re: Meeting? Time Y Z m 1 receive 5 11 Re: Meeting? m 2 m 3 Meeting? Re: Meeting? 31

32 Next time A brief introduction to Go (by Zach) Before Take the Go tour if you haven t! Join Piazza Look for a team 32

Distributed Systems. To do. q Distributed systems q This class q Models and architectures q Next: A brief introduction to Go

Distributed Systems. To do. q Distributed systems q This class q Models and architectures q Next: A brief introduction to Go Distributed Systems To do q Distributed systems q This class q Models and architectures q Next: A brief introduction to Go Image from http://www.vs.inf.ethz.ch/about/zeit.jpg Welcome! What is this all