X10: Performance and Productivity at Scale

1. X10:Performance and Productivity at Scale 李强

2. X10 Project status • X10 is developed by the IBM PERCS project as part of the DARPA program on High Productivity Computing Systems (HPCS) • Target markets: Scientific computing, business analytics • X10 2.4.2 release in Feb 2014 • Java-backend • C++-backend

3. X10 overview • X10 is an instance of the Asynchronous PGAS model in the Java family • Threads can be dynamically created under programmer control (as opposed to SPMD execution of MPI, UPC, FORTRAN) • n distinct threads, p distinct memories (n <> p) • APGAS model, the Asynchronous, Partitioned Global Address Space model.

4. X10 Hello world! class Generic datatype method • 1 // file HelloWorld.x10 • 2 public class HelloWorld { • 3 public static def main(args: Array[String](1)):Void { • 4 x10.io.Console.OUT.println("Hello, World"); • 5 } • 6 } package

5. X10 Basics • X10 is an object-oriented language based on Java • Base data types • Non-numeric: Boolean, Byte, Char and String • Fixed point: Short, Int and Long • Floating point: Float, Double and Complex • Top level containers: classes and interfaces, grouped into packages • Objects are instantiated from classes

6. Difference with java • var n:Long = 0; • public static val solution = [1,2,3,4,5] ; • val N:long; • Fields may be mutable (var) or immutable (val). • The type of a mutable field must always be specified. • A mutable field may or may not be initialized. • The type of an immutable field may be omitted if the field declaration specifies an initializer. • Function as argument, x10 supports both OO and functional programming.

7. X10 Parallelism • Parallelism = Activities + Places • Basic parallel constructs (async, at, finish, atomic) • Activities • All X10 programs begin with a single activity executing main in place 0 • Places hold activities and objects

8. public class TutAtomic2 { constint a = new boxedInt(100); constint b = new boxedInt(100); public static atomic void incr_a() { a.val++ ; b.val-- ; } public static atomic void decr_a() { a.val-- ; b.val++ ; } public static void main(String args[]) { int sum; finish { async for (inti=1 ; i<=10 ; i++ ) incr_a(); for (inti=1 ; i<=10 ; i++ ) decr_a(); } atomic sum = a.val + b.val; System.out.println("a+b = " + sum); } // main() } // TutAtomic2

9. Limitations of using a Single Place • Largest deployment granularity for a single place is a single SMP • Smallest granularity can be a single CPU or even a single hardware thread • Single SMP is inadequate for solving problems with large memory and compute requirements • X10 solution: incorporate multiple places as a core foundation of the X10 programming model  Enable deployment on large-scale clustered machines, with integrated support for intra-place parallelism

10. Scalable X10: using multiple places • class HelloWholeWorld { • public static def main(args:Array[String](1)):void { • for (var i:Int=0; i<Place.MAX_PLACES; i++) { • valiVal = i; • async at (Place.places(iVal)) { • Console.OUT.println("Hello World from place "+here.id); • } • } • } • } Hello World from place 0 Hello World from place 2 Hello World from place 3 Hello World from place 1

11. at – place shift • Shift current activity to a place to evaluate an expression, then return • Copy necessary values from calling place to callee place, discard when done

12. A more complicated example

13. A more complicated example

14. Thanks for your attention!