X10 X10 Jonathan Lee Jonathan Lee Daniel Lee Daniel Lee What is - - PowerPoint PPT Presentation
X10 X10 Jonathan Lee Jonathan Lee Daniel Lee Daniel Lee What is X10? What is X10? Programming language designed for high- Programming language designed for high- performance, high-productivity performance, high-productivity
Programming language designed for high-
Development at IBM Research
Object oriented (OO) Language
Intended to have simple and clear
Introduce a new programming language
Use the Java programming language as a
Added a
Added a few new things, took away some old few new things, took away some old things things
Uses partitioned global address space
Collection of data objects and activities (think of
Collection of data objects and activities (think of as threads) that operate on the data as threads) that operate on the data
Can think of as a
Can think of as a “ “virtual shared-memory multi- virtual shared-memory multi- processor processor” ”
Every X10 activity runs in a place
Every X10 activity runs in a place
Can get reference to the current place with the
Can get reference to the current place with the constant constant here
here
Places are ordered and the methods
Places are ordered and the methods next()
next() and
and
prev prev() () can be used to cycle through them
can be used to cycle through them
X10 uses PGAS (Partitioned Global
Each place has
Scalar objects are allocated completely at a
Elements of an array may be distributed
X10 activities, places, and PGAS
Can create asynchronous activities using
async
async (P) S (P) S
Spawns an activity at the place designated by
Spawns an activity at the place designated by P to execute S P to execute S
Creates parallelism!
Can be thought of as extremely
System.out. System.out.println println(1); (1); async async (place.next()) { (place.next()) { System.out. System.out.println println(2); (2); } } System.out. System.out.println println(3); (3);
Regions:
Simple contiguous ranges: [0:N]
Simple contiguous ranges: [0:N]
Multidimensional blocks: [0:N,0:M]
Multidimensional blocks: [0:N,0:M]
Can create arbitrary regions of any dimension
Can create arbitrary regions of any dimension
Region Operations:
Union:
Union: R1 || R2
R1 || R2
Intersection:
Intersection: R1 && R2
R1 && R2
Set Difference:
Set Difference: R1 - R2
R1 - R2
Distributions: Maps
Built in
Built in Distributions: Distributions:
Constant:
Constant: all points map to a single place all points map to a single place
Block: contiguous sets of
Block: contiguous sets of points equally divided points equally divided among places among places
Cyclic: Every Nth point
Cyclic: Every Nth point assigned to a place assigned to a place
Distribution Operations:
Distribution Operations:
Also include:
Also include:
Range Restriction:
Range Restriction: D | R D | R
Place Restriction:
Place Restriction: D | D | P P
Indexing for places:
Indexing for places: D[p] D[p] Example: Block Star Distribution
Example: Block Star Distribution
Distribution d = dist.factory.block([0,N],places); Distribution d = dist.factory.block([0,N],places); Distribution Distribution blockstar blockstar = [0:-1,0:-1]->here; = [0:-1,0:-1]->here; for (point p : d) { for (point p : d) { blockstar blockstar = = blockstar blockstar || [0:M]->d[i]; || [0:M]->d[i]; } }
X10 Arrays:
X10 Arrays:
Takes a distribution as a parameter
Takes a distribution as a parameter to assign data to to assign data to places places
Example:
Example: double[.]
double[.] data = new double[[0:N]->here]; data = new double[[0:N]->here];
Built in and user defined functions
Built in and user defined functions support support
Scans
Scans
Overlays
Overlays
Reductions
Reductions
Lifting
Lifting
Initialization
Initialization
for (point p : R) S
for (point p : R) S
Pointwise
Pointwise for for sequential iteration by a single for for sequential iteration by a single activity activity
Equivalent to
Equivalent to Java Java foreach foreach loops loops
ν ν Example:
Example:
Region r = Region r = [0:N]; [0:N]; int int[.] x = new [.] x = new int int[ [r- r->here]; >here]; for (point for (point p(i) : r) { p(i) : r) { x[p] = x[p] = i * 2; i * 2; } }
foreach
foreach (point p : R) S (point p : R) S
For parallel iteration in a single place
For parallel iteration in a single place
ν ν ≡
≡ for (point p : R)
for (point p : R) async async (here) { S } (here) { S }
ν ν Example:
Example:
Region r = Region r = [0:N]; [0:N]; int int[.] x = new [.] x = new int int[ [r- r->here]; >here]; foreach foreach (point (point p(i) : r) { p(i) : r) { x[p] = x[p] = i * 2; i * 2; } }
ateach
ateach (point p : D) S (point p : D) S
For parallel iteration across multiple places
For parallel iteration across multiple places
ν ν ≡
≡ for (point p : D)
for (point p : D) async async (D[p]) { S } (D[p]) { S }
ν ν Example:
Example:
Distribution d = Distribution d = [0:4]->place(0) || [0:4]->place(0) || [5:9]->place(1); [5:9]->place(1); int int[.] x = new [.] x = new int int[d]; [d]; ateach ateach (point (point p(i) : r) { p(i) : r) { x[p] = x[p] = i * 2; i * 2; } }
f = future(P) E f = future(P) E
Spawns an activity at place P to execute expression E
Spawns an activity at place P to execute expression E
When
When parent activity wants the result of E, it executes parent activity wants the result of E, it executes a a f.force()
f.force()
Parent activity blocks until the
Parent activity blocks until the future activity completes future activity completes
ν ν Example:
Example:
Distribution d = [0:4]->place(0) || Distribution d = [0:4]->place(0) || [5:9]->place(1); [5:9]->place(1); int int[.] x = new [.] x = new int int[d] (point (i)) { return [d] (point (i)) { return i; }; i; }; Future< Future<int int> fx5 = future (place(1)) { > fx5 = future (place(1)) { x[5] }; x[5] }; … … int int x5 = fx5.force(); x5 = fx5.force();
X10
X10’ ’s synchronization mechanism s synchronization mechanism
Acts much like a barrier
Acts much like a barrier
Activities register with a clock
Activities register with a clock
An activity can perform a
An activity can perform a next next operation to
indicate that it is ready to advance all the clocks indicate that it is ready to advance all the clocks it is registered with it is registered with
When all activities registered with clock perform
When all activities registered with clock perform next command, activities on clock can next command, activities on clock can continue continue
finish S
Essentially a join
Essentially a join
Must block until
Must block until all child all child activities recursively activities recursively complete complete
Also acts as
Also acts as aggregation point for aggregation point for exceptions exceptions
Example:
System.out. System.out.println println( (“ “start start” ”); ); finish finish foreach foreach(point (i,j) : [0:N,0:M]) { (point (i,j) : [0:N,0:M]) { System.out. System.out.println println(N * (N * i + j); i + j); } } System.out. System.out.println println( (“ “end end” ”); );
atomic S
atomic S
Such a statement is executed by the activity as
Such a statement is executed by the activity as if in a single step during which all other if in a single step during which all other activities are frozen activities are frozen
Type system ensures that statement S will
Type system ensures that statement S will dynamically access only local data dynamically access only local data
Conditional atomic
when(e) { s }
when(e) { s }
await(e)
await(e)
Uses polyglot to generate Java code
Leverages
Leverages java threads to achieve concurrence, java threads to achieve concurrence, but not much place partitioning but not much place partitioning
Runtime big and fat; lots of checks and
Runtime big and fat; lots of checks and indirection indirection
Compiler is
Compiler is fairly simplistic fairly simplistic
Java syntax and libraries
Constructs
Easy to use some constructs to gain some
Hard to load balance places
Hard to load balance places
Implementation is slow and compiler is
Implementation is slow and compiler is simplistic simplistic
Since implementation uses inner classes, final
Since implementation uses inner classes, final modifiers need to be added in some places modifiers need to be added in some places
At current state,
At current state, using parallelism constructs using parallelism constructs aggressively aggressively is slower is slower
Crypto
Jacobi
Questions?