Presentation is loading. Please wait.

Presentation is loading. Please wait.

Large-Scale Distributed Systems Andrew Whitaker CSE451.

Similar presentations


Presentation on theme: "Large-Scale Distributed Systems Andrew Whitaker CSE451."— Presentation transcript:

1 Large-Scale Distributed Systems Andrew Whitaker CSE451

2 Textbook Definition “A distributed system is a collection of loosely coupled processors interconnected by a communication network” Typically, the nodes run software to create an application/service  e.g., 1000s of Google nodes work together to build a search engine

3 Why Not to Build a Distributed System (1) Must handle partial failures  System must stay up, even when individual components fail Amazon.com

4 Why Not to Build a Distributed System (2) No global state  Machines can only communicate with messages This makes it difficult to agree on anything  “What time is it?”  “Which happened first, A or B?” Theory: consensus is slow and doesn’t work in the presence of failure  So, we try to avoid needing to agree in the first place A B

5 Reasons to Build a Distributed System (1) The application or service is inherently distributed Andrew Whitaker Joan Whitaker

6 Reason to Build a Distributed System (2) Application requirements  Must scale to millions of requests / sec  Must be available despite component failures This is why Amazon, Google, Ebay, etc. are all large distributed systems

7 Internet Service Requirements Basic goal: build a site that satisfies every user requests Detailed requirements:  Handle billions of transactions per day  Be available 24/7  Handle load spikes that are 10x normal capacity  Do it with a random selection of mismatched hardware

8 An Overview of HotMail (Jim Gray) ~7,000 servers 100 backend stores with 300TB (cooked) Many data centers Links to  Internet Mail gateways  Ad-rotator  Passport ~ 5 B messages per day 350M mailboxes, 250M active ~1M new per day. New software every 3 months (small changes weekly).

9 Availability Strategy #1: Perfect Hardware Pay extra $$$ for components that do not fail People have tried this  “fault tolerant computing” This isn’t practical for Amazon / Google:  It’s impossible to get rid of all faults  Software and administrative errors still exist

10 Availability Strategy #2: Over- provision Step 1: buy enough hardware to handle your workload Step 2: buy more hardware Replicate

11 Benefits of Replication Scalability Guards against hardware failures Guards against software failures (bugs)

12 Replication Meets Probability p is probability that a single machine fails Probability of N failures is: 1-p^n Site unavailability

13 Availability in the Real World Phone network: 5 9’s  99.999% available ATMs: 4 9’s  99.99% available What about Internet services?  Not very good…

14 2006: typical 97.48% Availability 97.48% Source: Jim Gray

15 Netcraft’s Crisis-of-the-Day

16 What Gives? Why isn’t simple redundancy enough to give very high availability?

17 Failure Modes Fail-stop failure: A component fails by stopping  It’s totally dead: doesn’t respond to input or output  Ideally, this happens fast Like a light-bulb Byzantine failure: Component fails in an arbitrary way  Produces unpredictable output

18 Byzantine Generals Basic goal: reach consensus in the presence of arbitrary failures Results:  More than 2/3 of the nodes must be “loyal” 3t + 1 nodes with t traitors  Consensus is possible, but expensive Lot’s of messages Many rounds of communication In practice, people assume that failures are fail- stop, and hope for the best…

19 Example of a non Fail-Stop Failure Server Load balancer Internet Load Balancer uses a “Least Connections” policy Server fails by returning an HTTP error 400 Net result: “failed” server becomes a black hole Amazon.com

20 Correlated Failures In practice, components often fail at the same time  Natural disasters  Security vulnerabilities  Correlated manufacturing defects  Human error…

21 Human error Human operator error is the leading cause of dependability problems in many domains Source: D. Patterson et al. Recovery Oriented Computing (ROC): Motivation, Definition, Techniques, and Case Studies, UC Berkeley Technical Report UCB//CSD-02-1175, March 2002. Public Switched Telephone Network Average of 3 Internet Sites Sources of Failure

22 Understanding Human Error Administrator actions tend to involve many nodes at once:  Upgrade from Apache 1.3 to Apache 2.0  Change the root DNS server  Network / router misconfiguration This can lead to (highly) correlated failures

23 Learning to Live with Failures If we can’t prevent failures outright, how can we make their impact less severe? Understanding availability:  MTTF: Mean-time-to-failure  MTTR: Mean-time-to-repair  Availability = MTTR / (MTTR + MTTF) Approximately MTTR / MTTF Note: recovery time is just as important as failure time!

24 Summary Large distributed systems are built from many flaky components  Key challenge: don’t let component failures become system failures Basic approach: throw lots of hardware at the problem; hope everything doesn’t fail at once  Try to decouple failures  Try to avoid single points-of-failure  Try to fail fast Availability is affected as much by recovery time as by error frequency


Download ppt "Large-Scale Distributed Systems Andrew Whitaker CSE451."

Similar presentations


Ads by Google