![16
Data Parallelism
def numbers = [10,6,3,8,1,5,7,2,4,9]
def sleeper = {number -> sleep number*100; print "${number}, "}
withPool(numbers.size()){
def sorter = sleeper.asyncFun()
numbers.each{sorter it}
}](https://image.slidesharecdn.com/itamartayer-gpars-141216152451-conversion-gate02/85/GPARS-Lessons-from-the-parallel-universe-Itamar-Tayer-CoolaData-16-320.jpg)
GPARS: Lessons from the parallel universe - Itamar Tayer, CoolaData
- 1. Lessons from the Parallel Universe November 2015 Itamar Tayer Otiot.org Itamar.tayer@cooladata.com GPARS @CoolaData LinkedIn.com/company/CoolaData www.facebook.com/Cooladata plus.google.com/+CooladataGplus/posts
- 2. AGENDA Enters Gpars Actors Try groovy! Agents Data parallelism & Fork/Join Data flow Parallel programming crisis What’s wrong with threads?
- 3. 3 Try Groovy • Almost zero learning curve for Java developers • Seamless integration with Java classes and libraries • Lots of good libraries • Grails • Gradle • Gorm • Spock • Json • And much more…
- 4. 4 Parallel Computing Crisis • Computational power is achieved by increasing the number of cores on the chip, rather than making each core faster • More and more models are using a (large) cluster of machines to solve complex problems • We need a computing paradigm that allow us to create reliable parallelism in a natural way
- 5. 5 What About 100 Cores?
- 6. 6
- 7. 7 What’s Wrong with Java Threads? • “Threads are evil” • Easy to get wrong, hard to get right • Hard to reason about the model state • Bugs (Dead locks, race conditions etc. ) are difficult to discover and track • Couples applicative and parallelism code • Limited, expensive and low level resources • Threading model is only applicable to local parallelism, not for distribution
- 8. 8 Pointing to the Problem… class Point { private int x,y; public setX(int newVal){ | //T1 x=newVal; | p.setX(2) } | p.setY(2) public setY(int newVal){ | //T2 y=newVal; | p.setX(4) } | p.setY(4) }
- 9. 9 Pointing to the Problem… class Point { private int x,y; public synchronized setX(int newVal){ x=newVal; } public synchronized setY(int newVal){ y=newVal; } }
- 10. 10 Pointing to the Problem… class Point { private int x,y; public synchronized set(int newX, int newY){ x=newX; y=newY; } }
- 11. 11 Pointing to the Problem… class Point { private int x,y; public synchronized set(Integer newX, Integer newY){ if (newX != null) x=newX; if (newY != null) y=newY; } }
- 12. 12 Another Observation public class ValueHolder { private List listeners = new LinkedList(); private int value; public interface Listener { public void valueChanged(int newValue); } public void addListener(Listener listener) { Listeners.add(listener); } public void setValue(int newValue) { value = newValue; Iterator i = listeners.iterator(); while(i.hasNext()) { ((Listener)i.next()).valueChanged(newValue); } } }
- 13. 13 We Need a Better Paradigm • Describe parallelism problems in an abstract, application agnostic model • Solve concurrency issues at the infrastructure level • Allow developer concentrate on application logic • Prevent developers from dealing with concurrency primitives • Just like memory is handled in JAVA!
- 14. 14 Enters GPARS • The groovy library for parallelism • Bundled in groovy since 1.8 • Offers various parallelism models • Clean and concise code due to groovy flexibility
- 15. 15 Data Parallelism def range = 1..100 println Runtime.getRuntime().availableProcessors() //sequential range.each { println it + ',' + Thread.currentThread()} //concurrent ParallelEnhancer.enhanceInstance(range) range.asConcurrent { range.each {println it + ',' + Thread.currentThread()} }
- 16. 16 Data Parallelism def numbers = [10,6,3,8,1,5,7,2,4,9] def sleeper = {number -> sleep number*100; print "${number}, "} withPool(numbers.size()){ def sorter = sleeper.asyncFun() numbers.each{sorter it} }
- 17. 17 Actor • Very common and reliable parallelism model • System is built from independent actors that communicate with each other via messages • Every actor has a ‘mail box’ for messages, and messages are processed one by one • A scheduler provide threads to actors that need to process a message • Program is naturally concurrent, no need to consider concurrency in application code
- 18. 18 Actor Model
- 19. ONLINE GAMING MODEL Actor Case Study 19 ZONE 01 ZONE 03 ZONE 02
- 20. 20 Groovy Actor final class MyDDA extends DynamicDispatchActor { def myState=0; void onMessage(String message) { myState+=1 println "Received string ${myState}" } void onMessage(Integer message) { myState+=2 println "Received integer ${myState}" }
- 21. 21 Using The Actor def myActor = new MyDDA().start() final Thread t1 = Thread.start {while (true) { myActor.send "Hello“ sleep 1000 } } final Thread t2 = Thread.start {while (true) { myActor << 12 sleep 1000 } }
- 23. 23 Agents • Sometimes sharing memory does make sense • Agent offers another level of indirection between a reference and it’s actual state • Agent encapsulates the execution Ref X X1 X2 Agent X
- 24. 24 Groovy Agent class Point { int x,y } agent = new Agent(new Point()) agent.send{it.x=2;it.y=2} agent<<{it.x=4;it.y=4} sleep(1000) def value = agent.val; println "$value.x“ println "$value.y";
- 25. 25 Data Flow • Data flow model allow us to divide our problem to smaller parallel computations • Each sub task is written independently, and the framework takes care for putting it all together • each data flow task can get a value only once during it’s life time. Reading the value will block until it’s ready • This will align all the tasks in the right order, and create a completely deterministic flow
- 26. 26 Data Flow final def x = new DataflowVariable() final def y = new DataflowVariable() final def z = new DataflowVariable() task { z << x.val + y.val } task { x << 10 //or get some value from the network } task { y << 5 //or do a long computation } println "Result: ${z.val}"
- 27. Build vs. Buy CoolaData Offers the Best of Both Worlds 27 ETL REAL TIME PROCESSING HBASE Couchbase CASSANDRA INTERACTIVE PROCESSING Exasol Vertica Redshift BATCH PROCESSING HADOOP HIVE REAL-TIME PROCESSING (STORM, KINSESIS) DATA VISUALIZATION (EXCEL, TABLEAU, QlikView) STRUCTURED AND UNSTRUCTURED DATA (HDFS, S3) vs • Lower cost of ownership • Faster time to market • Future proof: customizable and open • Stronger analytical power enabling predictive and proactive analytics • Development requires specific know-how and experience, not core to the business
- 28. 28 Low Risk, Quick 2 Results Grows with your organizational needs Proven