Optimistic Parallelism Requires Abstractions Milind Kulkarni, - - PowerPoint PPT Presentation

Optimistic Parallelism Requires Abstractions

Milind Kulkarni, Keshav Pingali – The University of Texas at Austin Bruce Walter, Ganesh Ramanarayanan, Kavita Bala and L. Paul Chew – Cornell University

Optimistic Parallelism Requires Abstractions

Milind Kulkarni, Keshav Pingali – The University of Texas at Austin Bruce Walter, Ganesh Ramanarayanan, Kavita Bala and L. Paul Chew – Cornell University

PLDI 2007 June 11th, 2007

Motivation

✦ Parallel programming very important ✦ Multicore processors ✦ Parallel programming is hard! ✦ Limited success in domains which deal with

structured data

✦ Array programs ✦ Database applications ✦ What about irregular applications which deal

with unstructured data?

✦ Compile time techniques have failed

3

PLDI 2007 June 11th, 2007

Galois System: Core Beliefs

✦ Irregular applications have worklist-style data

parallelism

✦ Optimistic parallelization is crucial ✦ Parallelism should be hidden within natural

syntactic constructs

✦ High level application semantics are critical

for parallelization

4

PLDI 2007 June 11th, 2007

Outline

✦ Two challenge problems ✦ Galois programming model and

implementation

✦ Evaluation ✦ Related Work ✦ Conclusions

5

PLDI 2007 June 11th, 2007

Delaunay Mesh Refinement

✦ Iterative refinement procedure to produce

guaranteed quality meshes

6

PLDI 2007 June 11th, 2007

Delaunay Pseudo-code

7

Mesh m = /* read in mesh */ WorkList wl; wl.add(mesh.badTriangles()); while (wl.size() != 0) { Element e = wl.get(); if (e no longer in mesh) continue; Cavity c = new Cavity(e); c.expand(); c.retriangulate(); mesh.update(c); wl.add(c.badTriangles()); }

PLDI 2007 June 11th, 2007

Delaunay Pseudo-code

8

Mesh m = /* read in mesh */ WorkList wl; wl.add(mesh.badTriangles()); while (wl.size() != 0) { Element e = wl.get(); if (e no longer in mesh) continue; Cavity c = new Cavity(e); c.expand(); c.retriangulate(); mesh.update(c); wl.add(c.badTriangles()); } Worklist idiom

PLDI 2007 June 11th, 2007

Finding Parallelism

✦ Can expand multiple cavities in parallel ✦ Provided cavities do not overlap ✦ Determining this statically is impossible ✦ Solution: Optimistic parallel execution

9

PLDI 2007 June 11th, 2007

Agglomerative Clustering

✦ Create binary tree of points in a space in

bottom-up fashion

✦ Always choose two closest points to cluster

10

a b c d e a b c d e a b c d e (a) Data points (b) Hierarchical clusters (c) Dendrogram

PLDI 2007 June 11th, 2007

Agglomerative Clustering

✦ Two key data structures ✦ Priority Queue – Keeps pairs of points

<p,n> where n is the nearest neighbor of p

✦ Ordered by distance ✦ KD-tree – Spatial structure to find nearest

neighbors

11

PLDI 2007 June 11th, 2007

Finding Parallelism

✦ Priority queue functions as a worklist ✦ Seems to be completely sequential ✦ If clusters are independent, can be done in

parallel

12 a b c d e

PLDI 2007 June 11th, 2007

Lessons Learned

✦ Worklist-style data parallelism ✦ May be dependences between iterations ✦ However, worklist abstractions are missing

from the code

✦ Concurrent access to shared objects a must ✦ worklist, priority queue, kd-tree

13

Galois Programming Model and Implementation

PLDI 2007 June 11th, 2007

Programming Model

✦ Object-based shared

memory model

✦ Client code must

invoke methods to access object state

✦ Client code has

sequential semantics

✦ But runtime system

may execute code in parallel

15

Client Code Galois Objects

PLDI 2007 June 11th, 2007

Worklist Abstractions

✦ Iterators over collections

✦ foreach e in set S do B(e)

✦ Iterations can execute in any order ✦ As in Delaunay mesh refinement

✦ foreach e in poSet S do B(e)

✦ Iterations must respect ordering of S ✦ As in agglomerative clustering ✦ May be dependences between iterations ✦ Sets can change during execution

16

PLDI 2007 June 11th, 2007

Delaunay Example

17

Mesh m = /* read in mesh */ WorkList wl; wl.add(mesh.badTriangles()); while (wl.size() != 0) { Element e = wl.get(); if (e no longer in mesh) continue; Cavity c = new Cavity(e); c.expand(); c.retriangulate(); mesh.update(c); wl.add(c.badTriangles()); }

PLDI 2007 June 11th, 2007

Delaunay Example

18

Mesh m = /* read in mesh */ WorkList wl; wl.add(mesh.badTriangles()); foreach Element e in wl { if (e no longer in mesh) continue; Cavity c = new Cavity(e); c.expand(); c.retriangulate(); mesh.update(c); wl.add(c.badTriangles()); } rest of code unchanged

PLDI 2007 June 11th, 2007

Delaunay Example

19

Mesh m = /* read in mesh */ WorkList wl; wl.add(mesh.badTriangles()); foreach Element e in wl { if (e no longer in mesh) continue; Cavity c = new Cavity(e); c.expand(); c.retriangulate(); mesh.update(c); wl.add(c.badTriangles()); }

Iterators expose worklist abstraction to runtime system

PLDI 2007 June 11th, 2007

Execution Model

✦ Master thread begins execution ✦ When it encounters an iterator, it uses helper

threads to aid in execution of iterations

✦ Iterations assigned to thread according to

scheduling policy (for now, dynamic to ensure load balance)

✦ Parallel execution of iterator must respect

sequential semantics of iterator

✦ Concurrent access control ✦ Serializability of iterations

20

PLDI 2007 June 11th, 2007

Concurrent Access

✦ Concurrent invocations

to a shared object must not interfere

✦ Our current

implementation uses locks

✦ Can use other

techniques such as TM

21

S.add(y) S.add(x)

S

PLDI 2007 June 11th, 2007

Serializability

22

S.contains?(x) S.remove(x) S.add(x)

S Workset

S.add() ... = S.get() S.add() ... = S.get()

(a) Interleaving is illegal (b) Interleaving is legal (and necessary)

PLDI 2007 June 11th, 2007

Semantic Commutativity

✦ Method calls which commute can be

interleaved

✦ Else, commutativity violation ✦ Property of abstract data type ✦ Implementation independent

23

PLDI 2007 June 11th, 2007

Galois Classes

✦ Inverse methods ✦ Allow for rollback

when commutativity violated

✦ Commutativity and

inverse specified through interface annotation

24

class SetInterface { void add(T x); [commutes] add(y) {y != x} remove(y) {y != x} contains(y) {y != x} [inverse] remove(x) bool contains(T x); [commutes] add(y) {y != x} remove(y) {y != x} ... }

PLDI 2007 June 11th, 2007

Galois Classes

✦ Inverse methods ✦ Allow for rollback

when commutativity violated

✦ Commutativity and

inverse specified through interface annotation

25

class SetInterface { void add(T x); [commutes] add(y) {y != x} remove(y) {y != x} contains(y) {y != x} [inverse] remove(x) bool contains(T x); [commutes] add(y) {y != x} remove(y) {y != x} ... }

Galois Classes expose abstractions to the runtime system

PLDI 2007 June 11th, 2007

Runtime System

✦ Two main components: ✦ Global commit pool ✦ Manages iterations ✦ Similar to reorder buffer in OOE

processors

✦ Per object conflict logs ✦ Detects commutativity violations ✦ Triggers aborts if commutativity violated

26

PLDI 2007 June 11th, 2007

Evaluation

✦ Evaluation platform: ✦ Implementation in C++ ✦ gcc compiler on Red Hat Linux ✦ 4 processor, shared memory system ✦ Itanium 2 @ 1.5 GHz

27

PLDI 2007 June 11th, 2007

Evaluation – Delaunay

✦ Three different versions of benchmark ✦ reference – purely sequential code ✦ FGL – hand-written, optimistic parallel code

using fine-grained locking

✦ meshgen – Galois version of code ✦ Input mesh generated using Triangle ✦ ~10K triangles ✦ ~4K bad triangles

28

PLDI 2007 June 11th, 2007

Abort Ratios

✦ Optimism must be warranted ✦ Conflicts lead to rollbacks, which waste

work

✦ FGL and meshgen have abort ratios <1% on 4

processors

✦ Closely tied to scheduling policy ✦ Choice of proper scheduling policy is

crucial for good performance

29

PLDI 2007 June 11th, 2007

1 2 3 4

# of processors

1 1.5 2 2.5 3

Speedup

reference FGL meshgen 1 2 3 4

# of processors

2 4 6 8

Execution Time (s)

reference FGL meshgen

Evaluation – Delaunay

30

PLDI 2007 June 11th, 2007

1 2 3 4

# of processors

1 1.5 2 2.5 3

Speedup

reference FGL meshgen 1 2 3 4

# of processors

2 4 6 8

Execution Time (s)

reference FGL meshgen

Evaluation – Delaunay

31

~3x speedup

PLDI 2007 June 11th, 2007

Performance Breakdown

32

Client Object Runtime 5 10 15 20 1 proc 4 proc

Cycle (billions)

13.8951 18.8501 5 10 15 20 1 proc 4 proc

Instructions (billions)

16.8889 17.4675

PLDI 2007 June 11th, 2007

Related Work

✦ Weihl, 1988 – Concurrency control using

commutativity properties of ADTs

✦ Rinard & Diniz, 1996 – Static commutativity

analysis for parallelization

✦ Wu & Padua, 1998 – Exploiting semantic

properties of containers in compilation

✦ Ni et al, 2007 – Open nesting using abstract

locks

33

PLDI 2007 June 11th, 2007

Conclusions

✦ Optimistic parallelism necessary to parallelize

irregular, worklist-based applications

✦ Need to exploit high-level semantics ✦ Iterators to expose parallelism ✦ Galois classes to expose semantics of

• bjects

34

Thank You!

Email: milind@cs.utexas.edu