Dynamic ABP Work Stealing

1
Dynamic ABP Work Stealing

• Presenter Yossi Lev
• Sun Microsystems Laboratories Tel Aviv
  University
• Joint work with
• Danny Hender Nir Shavit
• Tel-Aviv University Sun Microsystems Laboratories
  Tel Aviv University

2
Work Stealing Algorithms

• Processes works with their local workpile.
• If emptied, a process steals item from another
  workpile.

3
ABP Work Stealing AlgorithmArora Blumofe
Plaxton 98

• A non-blocking work stealing algorithm
• Intended for multiprogrammed environments.
• The local workpile is implemented as an
  array-based deque.
• Operations by a process on its local deque
  require only read writes unless there is only
  one element.

4
The Problem

• Arrays may overflow.
• Solutions people used
• Exploit the array space better Reset-On-Empty,
  Cyclic Deque.
• Specialized overflow handling Parallel Garbage
  CollectionDetlefs, Flood, Shavit, Zhang

5
The ABP Deque
6
Dynamic ABP Deques

• Our Solution Implement the deque by a doubly
  linked list of small arrays. Save available nodes
  in a shared pool.

The Challenge Keep the good performance of the
original ABP deque.
Top
Bottom
Shared Pool of Nodes
Bottom
Top
Top
Bottom
7
Challenges

• Empty Deque Detection
• Maintaining the List Structure
• Choosing the Node Size.

8
Challenge I Empty Deque

• Empty deque detection Checking for crossing
  situation also between nodes.

9
Challenge II List Maintenance

• Next and previous pointers are updated only by
  local process (keep good performance).
• Deal with Loops.

How to distinguish between the two when checking
emptiness?
Still point to Node B
10
Challenge III Node Size

• What size should the nodes be?
• Use a local base-node with a big array for most
  of the deques entries.
• Use share nodes with small arrays for the rest.

11
Reset-On-Empty

• Support the Reset-On-Empty heuristic
• Especially important with large base-node.
• May need to change the target node on the fly.
• New empty scenarios arises.

Updating Top Scenario A CAS succeeds
Updating Top Scenario B CAS failed
12
Performance

• The Hood Library.
• With regular and dynamic ABP deques.
• Barnes Hut benchmark.
• Stand-Alone and Multiprogrammed Env.
• Setup
• The same total amount of memory for
  regular/dynamic versions

13
Performance (Cont)

• Stand alone run Dynamic ABP shows no performance
  overhead

• Multiprogrammed environment Dynamic shows
  higher stability.

14
Summary

• Array-based ABP deques can overflow. Therefore
  users must devise specialized overflow handling
  mechanisms.
• The new dynamic deque implementation eliminates
  this need.