2. OUTLINE
• CILK and CILK++ Language Features and
Usages
• Work stealing runtime
• CILK++ Reducers
• Conclusions
2
3. IDEALIZED SHARED
MEMORY ARCHITECTURE
3
• Hardware model
• Processors
• Shared global
memory
• Software model
• Threads
• Shared variables
• Communication
• Synchronization
Slide from Comp 422 Rice University Lecture 4
4. CILK AND CILK++
DESIGN GOALS
• Programmer friendly
• Dynamic tasking
• Parallel extension to C
• Scalable performance
• Efficient runtime system
• Minimum program overhead
4
5. CILK KEYWORDS
• Cilk: a Cilk function
• Spawn: call can execute asynchronously
in a concurrent thread
• Sync: current thread waits for all locally-
spawned functions
5
6. CILK EXAMPLE
cilk int fib(n) {
if (n < 2)
return n;
else {
int n1, n2;
n1 = spawn fib(n-1);
n2 = spawn fib(n-2);
sync;
return (n1 + n2);
}
}
6
Borrowed from Comp 422 Rice University Lecture 4
7. CILK++ EXAMPLE
int fib(n) {
if (n < 2)
return n;
else {
int n1, n2;
n1 = cilk_spawn fib(n-1);
n2 = fib(n-2);
cilk_sync;
return (n1 + n2);
}
}
7
Borrowed from Comp 422 Rice University Lecture 4
8. CILK++ EXAMPLE
WITH DAG
8
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
9. OUTLINE
• CILK and CILK++ Language Features and
Usages
• Work stealing runtime
• CILK++ Reducers
• Conclusions
9
10. WORK FIRST
PRINCIPLE
• Work: T1
• Critical path length: T∞
• Number of processor: P
• Expected time
• Tp = T1/P + O(T∞)
• Parallel slackness assumption
• T1/P >> C∞T∞
10
11. WORK FIRST
PRINCIPLE
• Minimize scheduling overhead borne by
work at the expense of increasing critical
path
• Tp ≤ C1Ts/P + C∞T∞
≈ C1Ts/P
Minimize C1 even at the expense of a larger
C∞
11
12. WORK STEALING
DESIGN GOALS
• Minimizing contentions
• Decentralized task deque
• Doubly linked deque
• Minimizing communication
• Steal work rather than push work
• Load balance across cores
• Lazy task creation
• Steal from the top of the deque
12
13. CILK WORK STEALING
SCHEDULER
13
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
14. CILK WORK STEALING
SCHEDULER
14
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
15. CILK WORK STEALING
SCHEDULER
15
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
16. CILK WORK STEALING
SCHEDULER
16
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
17. CILK WORK STEALING
SCHEDULER
17
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
18. CILK WORK STEALING
SCHEDULER
18
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
19. CILK WORK STEALING
SCHEDULER
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
20. CILK WORK STEALING
SCHEDULER
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
21. CILK WORK STEALING
SCHEDULER
21
Pictures from “Reducers and Other CILK+ HyperObjects”
Talk by Matteo Frigo (Intel). Pablo Halpern ( Intel).
Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
22. CILK WORK STEALING
SCHEDULER
22
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
23. CILK WORK STEALING
SCHEDULER
23
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
24. CILK WORK STEALING
SCHEDULER
24
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
25. CILK WORK STEALING
SCHEDULER
25
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
26. CILK WORK STEALING
SCHEDULER
26
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
27. TWO CLONE
STRATEGY
• Fast clone
• Identical in most respects to the C elision of the Cilk
program
• Very little execution overhead
• Sync statements compile to no op
• Allocates an continuation
• Program variables and instruction pointer
• Slow clone
• Convert a spawn schedule to slow clone only when it
is stolen
• Restores program state from activation frame that
contains local variables, program counter and other
parts of the procedure instance
27
29. SLOW CLONE
Slow_fib(frame * _cilk_frame){
restore states of the program
switch (_cilk_frame->header.entry)
{
fast_fib(_cilk_frame->n - 1 );
case 1: goto _cilk_sync1;
fast_fib(_cilk_frame->n - 2 );
case 2: goto _cilk_sync2;
sync (not a no op)
case 3: goto _cilk_sync3;
}
}
29
31. FRAMES
• C++ Main Frame
• Local variables of the procedure instance
• Temporary variables
• Linkage information for return values
31
32. FRAMES
• CILK++ Stack Frame
• Everything in C++ Main Frame
• Continuation
• Parent pointer
• Have exactly one child
• Used by Fast Clone
• A worker can have multiple Stack Frames
32
33. FRAMES
• CILK++ Full Frame (used by slow clone)
• Everything in CILK++ Stack Frame
• Lock
• Join counter
• List of children (has more than one
children)
• A worker has at most one Full Frame
33
34. FUNCTION CALL
34
Stack frame
Full frame
Extended Deque (Before Function Call)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
35. FUNCTION CALL
35
Stack frame
Full frame
Extended Deque (After Function Call)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
New stack
frame
36. SPAWN
36
Stack frame
Full frame
Extended Deque (Before Spawn Call)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
37. SPAWN
37
Stack frame
Full frame
Extended Deque (After Spawn Call)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
Set
continuation
in last stack
frame
38. RESUME FULL FRAME
38
Stack frame
Full frame
Extended DequeFunction call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
Set the full frame to be the only frame in the
call stack, resume execution on the
continuation
39. RANDOMLY STEAL
39
Stack frame
Full frame
Extended DequeFunction call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
Steal this call stack
40. RANDOMLY STEAL
40
Stack frame
Full frame
Extended DequeFunction call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
Steal this call stack
1 1 1
41. RANDOMLY STEAL
41
Stack frame
Full frame
Extended Deque
Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
1
1 1
42. PROVABLY GOOD
STEAL
42
Stack frame
Full frame
Extended DequeFunction call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
0
44. FUNCTION CALL
RETURN
44
Stack frame
Full frame
Extended Deque (Before Return from a Call Case1)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
45. FUNCTION CALL
RETURN
45
Stack frame
Full frame
Extended Deque (Return from a Call Case 1)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
46. FUNCTION CALL
RETURN
46
Stack frame
Full frame
Extended Deque (Return from a Call Case2)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
Worker executes an
unconditional steal
47. SPAWN RETURN
47
Stack frame
Full frame
Extended Deque (Before Spawn return Case 1)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
48. SPAWN RETURN
48
Stack frame
Full frame
Extended Deque (After Spawn return Case 1)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
49. SPAWN RETURN
49
Stack frame
Full frame
Extended Deque (Return from a SpawnCase2)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
Worker executes an
provably good steal
50. SYNC
50
Stack frame
Full frame
Extended Deque (Sync Case 1)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
Do nothing if it
is a stack
frame (No Op)
51. SYNC
51
Stack frame
Full frame
Extended Deque (Sync Case 2)Function call
Spawn
Call return
Spawn return
Sync
Randomly steal
Provably good
steal
Unconditionally
steal
Resume full
frame
Pop the frame,
provably good steal
52. OUTLINE
• CILK and CILK++ Language Features and
Usages
• Work stealing runtime
• CILK++ Reducers
• Conclusions
52
53. PROBLEMS WITH
NON-LOCAL VARIABLES
bool has_property(Node *)
List<Node *> output_list;
void walk(Node *x)
{
if (x) {
if (has_property(x))
output_list.push_back(x);
cilk_spawn walk(x->left);
walk(x->right);
cilk_sync;
}
}
53
54. REDUCER
DESIGN GOALS
• Support parallelization of programs
containing global variables
• Enable efficient parallel scaling by
avoiding a single point of contention
• Provide deterministic result for
associative reduce operations
• Operate independently of any control
constructs
54
55. REDUCER EXAMPLE
bool has_property(Node *)
List_append_reducer<Node *> output_list;
void walk(Node *x)
{
if (x) {
if (has_property(x))
output_list.push_back(x);
cilk_spawn walk(x->left);
walk(x->right);
cilk_sync;
}
}
55
56. HYPER OBJECTS
56
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
57. REDUCER
57
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
58. SEMANTICS OF
REDUCERS
• The child strand owns the view owned by
parent function before cilk_spawn
• The parent strand owns a new view,
initialized to identity view e,
• A special optimization ensures that if a
view is unchanged when combined with
the identity view
• Parent strand P own the view from
completed child strands
58
59. REDUCING OVER LIST
CONCATENATION
59
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
60. REDUCING OVER LIST
CONCATENATION
60
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
61. IMPLEMENTATION OF
REDUCER
• Each worker maintains a hypermap
• Hypermap
• Maps reducers to the views
• User
• The view of the current procedure
• Children
• The view of the children procedures
• Right
• The view of right sibling
• Identity
• The default value of a view
61
63. HYPERMAP CREATION
64
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
64. HYPERMAP CREATION
65
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
65. HYPERMAP CREATION
66
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
66. HYPERMAP CREATION
67
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
67. HYPERMAP CREATION
68
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
68. LOOK UP FAILURE
• Inserts a view containing an identity
element for the reducer into the
hypermap.
• Following the lazy principle
• Look up returns the newly inserted
identity view
69
69. RANDOM WORK
STEALING
A random steal operation steals a full frame
P and replaces it with a new full frame C in
the victim.
USERC ← USERP;
U S E R P ← 0/ ;
CHILDRENP←0/;
RIGHTP←0/.
70
70. RANDOM WORK
STEALING
71
Pictures from “Reducers and Other CILK+ HyperObjects” Talk by Matteo Frigo (Intel). Pablo
Halpern ( Intel). Charles E. Leiserson (MIT). Stephen Lewin-Berlin (Intel).
71. RETURN FROM A CALL
Let C be a child frame of the parent frame P
that originally called C, and suppose that C
returns.
• If C is a stack frame, do nothing,
• If C is a full frame.
• Transfer ownership of view
• Children and Right are empty
• USERP ← USERC
77
72. RETURN FROM A
SPAWN
Let C be a child frame of the parent frame P that
originally spawned C, and suppose that C returns.
• Always do USERC ← REDUCE(USERC,RIGHTC)
• If C is a stack frame, do nothing
• If C is a full frame
• If C has siblings,
• RIGHTL ← REDUCE(RIGHTL,USERC)
• C is the leftmost child
• CHILDRENP ←
REDUCE(CHILDRENP,USERC)
78
73. SYNC
A cilk_sync statement waits until all children have com-
pleted. When frame P executes a cilk_sync, one of following
two cases applies:
• If P is a stack frame, do nothing.
• If P is a full frame,
• USERP ← REDUCE(CHILDRENP,USERP).
82
75. OUTLINE
• CILK and CILK++ Language Features and
Usages
• Work stealing runtime
• CILK++ Reducers
• Conclusions
84
76. CONCLUSIONS
• CILK and CILK++ provide a programmer
friendly programming model
• Extension to C
• Incremental parallelism
• Scaling on future machines
• Non-compromising performance
• Work stealing runtime
• Minimizing overheads
• Reducers
85
77. FINAL NOTES
• Designed for an idealized shared memory
model
• Today’s architectures are typically NUMA
• Task creation can be lazier
• http://ieeexplore.ieee.org/xpls/abs_all.jsp?
arnumber=6012915&tag=1
• Cilk_for
• Divide and conquer parallelization
86
Notas del editor
CILK and CILK++ adopt the shared memory model, No uniform address, sockets, abstraction
If you have taken Comp 322, Spawn is very similar to the “async” keyword in Habanero Java, the Sync keyword is similar to the “finish” scopeCilk++ extends C++??
An example of thefibonacci sequence computation in cilk, Spawn two threads at each invocation of the function, notice the cilk keyword is used to denote a cilk function,
Cilk++ took away the cilk keyword, prefixed cilk_ to spawn and sync
Directed acyclic graphSpawn creates parallel executions, B and C, they join together and recombine to execute D
Work:The time needed to execute the program serialyParallel slackness assumption: number of processors is much smaller than average degree of parallelism
To support dynamic task creation
The Cilk runtime uses a specialworkstealing scheduler, There are two kinds of schedulers, worksharing, where all the workers steal from a unified task queue, it is less efficient for a number of reasons, There is a single lock potentially on the task queue to deal with contentionsThe queue could be empty, but there are still work leftThe workstealing runtime solvees the problem by building an extended deque for each worker, when a worker is out of work, it steals randomly from other workersWe will demonstrate the process in the next few slidesDecentralized Push work rather than pull work (when necessary)Loop contians a spawn, package child task, stack, single processor 9LAZY TASK CREATION
Steal from the top to reduce contentionSteal from the top to get bigger subtree (divide and conquer), larger task granularity, minimize stealsSteal from the top increase possible locality of the program (cache locality
The reason
All sync statements compile to no-ops because a fast clone never has any children when it is executing, we know at compile time that all previously spawned procedures have completed. Thus, no operations are required for a sync statementBefore it recursively spawns,
Looks a lot like orginal fib (highlight the original sequential code), the rest is bookeepingLittle bit bookeeping, Sig is the signature , included the pointer to the slow clone rountine, fibsig represents the slow cloneEntry point, instruction pointerComes back to the principle we described earlier
Uses fast_fib locally
Set continuation in original proc’s stack frameAllocates a stack frame for BPushes B’s stack frame to the tail of deque
Pick a random victim v, where v ̸= w. Repeat this step while the deque of v is empty. Remove the oldest call stack from the deque of v, and pro- mote all stack frames to full frames. For every promoted frame, increment the join counter of the parent frame (full by Invariant 3). Make every newly created child the right- most child of its parent. Let loot be the youngest frame that was stolen. Promote the oldest frame now in v’s extended deque to a full frame and make it the rightmost child of loot. Increment loot’s join counter. Execute a resume-full-frame action on loot.
Pick a random victim v, where v ̸= w. Repeat this step while the deque of v is empty. Remove the oldest call stack from the deque of v, and pro- mote all stack frames to full frames. For every promoted frame, increment the join counter of the parent frame (full by Invariant 3). Make every newly created child the right- most child of its parent. Let loot be the youngest frame that was stolen. Promote the oldest frame now in v’s extended deque to a full frame and make it the rightmost child of loot. Increment loot’s join counter. Execute a resume-full-frame action on loot.
Pick a random victim v, where v ̸= w. Repeat this step while the deque of v is empty. Remove the oldest call stack from the deque of v, and pro- mote all stack frames to full frames. For every promoted frame, increment the join counter of the parent frame (full by Invariant 3). Make every newly created child the right- most child of its parent. Let loot be the youngest frame that was stolen. Promote the oldest frame now in v’s extended deque to a full frame and make it the rightmost child of loot. Increment loot’s join counter. Execute a resume-full-frame action on loot.
Joint counter, frames left in heap, (0)Assert that the frame A begin stolen is a full frame and the extended deque is empty. Decrement the join counter of A. If the join counter is 0 and no worker is working on A, execute a resume-full-frame action on A. Otherwise, begin random work stealing.3
Assert that the frame A being stolen is a full frame, the extended deque is empty, and A’s join counter is positive. Decrement the join counter of A. Execute a resume-full- frame action on A.
Set continuation in original proc’s stack frameAllocates a stack frame for BPushes B’s stack frame to the tail of deque
Just removing a stack frame
This case the full frame has finished execution
Set continuation in original proc’s stack frameAllocates a stack frame for BPushes B’s stack frame to the tail of deque
Set continuation in original proc’s stack frameAllocates a stack frame for BPushes B’s stack frame to the tail of deque
This case the full frame has finished execution
Do nothing if it is a stack frame
Do nothing if it is a stack frame
Little modificationsDeterministic output even in the presence of output (associative)
Can be used to parallelize many programs containing global (or nonlocal) variables without locking, atomic updating, or the need to logically restructure the codeThe programmer can count on a deterministic result as long as the reducer operator is associative. Commutability is not requiredReducers opeerateindependenly of any control constructs, such as parallel for, and of any data structures that contribute their values to the final result
Little modificationsDeterministic output even in the presence of output (associative)
Fast clone uses identity view
Example of serial execution
Children of A would be {B, C}Right Sibling of B would be CUser would be view in A,
We distinguish two cases: the “fast path” when C is a stack frame, and the “slow path” when C is a full framebecause both P and C share the view stored in the map at the head of the deque to which both P and C belong. which transfers ownership of child views to the parent. The other two hypermaps of C are guaranteed to be empty and do not participate in the update
Set continuation in original proc’s stack frameAllocates a stack frame for BPushes B’s stack frame to the tail of deque
Just removing a stack frame
We distinguish two cases: the “fast path” when C is a stack frame, and the “slow path” when C is a full framebecause both P and C share the view stored in the map at the head of the deque to which both P and C belong. which transfers ownership of child views to the parent. The other two hypermaps of C are guaranteed to be empty and do not participate in the update
This case the full frame has finished execution
We distinguish two cases: the “fast path” when C is a stack frame, and the “slow path” when C is a full framebecause both P and C share the view stored in the map at the head of the deque to which both P and C belong. which transfers ownership of child views to the parent. The other two hypermaps of C are guaranteed to be empty and do not participate in the update
Again we distinguish the “fast path” when C is a stack frame from the “slow path” when C is a full frame:
If proc B finishes first,
If proc B finishes first, the results would be in children of A, If C finishes, it would be the left most, Children of A would just be a union of current children of A and UserCTwo of them are leftmost case
When C finishesC has a right sibling, B, so the result of C is accumulated into Right BWhen B finishes, the children of A has UserB
1. Doing nothing is correct because all children of P, if any exist, were stack frames, and thus they transferred ownership of their views to P when they completed. Thus, no outstanding child views exist that must be reduced into P’s. 2. Then after P passes the cilk_sync state- ment but before executing any client code, we perform the update. This up- date reduces all reducers of completed children into the parent.
Comparing reducers against mutual exclusion
Future scaling with dynmiac parallelismProvides a simple way to add incremental parallelismIncremental parallelization of programsInspired many future works, such as Habanero Java, Habanero C, X10,
Eagerly saving all the state, gather the states using an Exception when they make a steal