SLIDE 1
Chalmers University of Technology
1
Chalmers University of Technology
2
Talk Outline
- Synchronization in shared memory
multiprocessors
- Non-blocking Queue
– ABA Problem – Performance issues
- Conclusions
A Simple, Fast and Scalable Non-Blocking Concurrent FIFO Queue for - - PowerPoint PPT Presentation
A Simple, Fast and Scalable Non-Blocking Concurrent FIFO Queue for - - PowerPoint PPT Presentation
Chalmers University of Technology A Simple, Fast and Scalable Non-Blocking Concurrent FIFO Queue for Shared Memory Multiprocessor Systems Philippas Tsigas Yi Zhang Department of Computing Science Chalmers University of Technology 1
SLIDE 2
SLIDE 3
Chalmers University of Technology
3
Mutual Exclusion
- Traditional way for synchronization.
- Performance degradation under high
contention.
– Network contention – Lock convoy
- Complex (pessimistic) scheduling
analysis
– Priority inversion – Deadlock
SLIDE 4
Chalmers University of Technology
4
SLIDE 5
Chalmers University of Technology
5
Non-blocking Synchronization
- An alternative approach for
synchronization
- Lock-free and Wait-free
- Better performance in multiprocessor
systems
– No lock convoy – No priority inversion – No deadlock
SLIDE 6
Chalmers University of Technology
6
Non-blocking Queue
Previous Results
- Designed for Asynchronous Shared
Memory Multiprocessors
- Previous work
– Lamport (1983) – … … – Michael and Scott (1998) – … …
SLIDE 7
Chalmers University of Technology
7
Non-blocking Queue
Our Results
- The new non-blocking queue
- utperforms the best known
alternative implementation.
+ the new solution to the ABA problem together with + the lazy pointer updating to improve performance + the algorithmic design of the queue as a cyclic array
SLIDE 8
Chalmers University of Technology
8
ABA or Pointer Recycling Problem
- Occurs when read-modify-write is used
in lock-free computing with CAS atomic primitive
- Drawback of the CAS atomic primitive
- A Lot of overhead introduced to solve
it
SLIDE 9
Chalmers University of Technology
9
The Specification of CAS
Boolean CAS(int *mem, register old, new) { temp = *mem; if (temp == old) { *mem = new; return (TRUE);} else return FALSE; }
SLIDE 10
Chalmers University of Technology
10
ABA Problem (Example)
- Array-based Queue (Enqueue)
1. Loop 2. head = Queue.head 3. ... ... 4. if CAS(Queue.array[head],NULL,data) 5. ... ...
- 6. End loop
SLIDE 11
Chalmers University of Technology
11
ABA Problem (Example)
- Array-based Queue (Dequeue)
1. Loop 2. tail = Queue.tail 3. ... ... 4. if CAS(Queue.array[tail],data,NULL) 5. ... ...
- 6. End loop
SLIDE 12
Chalmers University of Technology
12
ABA Problem (Example)
Execution History (Empty Queue)
- P1
Enqueue 2 (Preempted) Enqueue 4
- P2
Enqueue ... ... Dequeue
SLIDE 13
Chalmers University of Technology
13
ABA Problem
Traditional Solution
- Using version number
– Splits word into two parts – Uses one part of the word as a version number – Increases the version number of the word whenever updating the word
- Not a complete solution, ABA can still
happen, when the version number runs
- ut of space.
SLIDE 14
Chalmers University of Technology
14
ABA Problem
Traditional Solution (Drawbacks)
- The actual pointer length is smaller
than the system pointer length
– Programmers must manage the pointer (memory) themselves – Limits the memory that can be accessed
- Tag operations introduce extra
- verhead
SLIDE 15
Chalmers University of Technology
15
ABA Problem
Our Solution
- Introduce a ghost copy of each value and
turn ABA to ABA’B´A
- For example, NULL means empty in the
Queue implementation
- Using NULL(0) and NULL(1) mean empty
cell
- Recycle the NULL values
- More NULL values can be introduced.
SLIDE 16
Chalmers University of Technology
16
Queue using Cyclical Array
SLIDE 17
Chalmers University of Technology
17
ABA Problem
Execution History (Empty Queue)
- P1
Enqueue 2 (Preempted) Enqueue 4
- P2
Enqueue at pos A ... ... Dequeue at pos A ... ... Enqueue at pos A ... ... Dequeue at pos A
SLIDE 18
Chalmers University of Technology
18
Performance Issues of Synchronization
- Network contention
– Access to shared memory – Spinning on shared memory – Cache coherent protocols
- Lock convoys
SLIDE 19
Chalmers University of Technology
19
Mutual Exclusion Based Solutions for Performance Issues
- Avoiding network contention
– Ticket lock – MCS Queue lock
- Avoiding lock-convoy effect
– Scheduler-conscious synchronization
SLIDE 20
Chalmers University of Technology
20
Non-blocking and Performance
- Avoiding the performance problems
- f lock-convoy from the beginning
- No much consideration about network
contention
SLIDE 21
Chalmers University of Technology
21
Observations on Queue
- perations
- CAS operations are network operations;
they generate a lot of traffic
- CAS operations are used when changing the
head and tail of the queue
- Head and tail of the queue do not have to
always point to the actual head and tail of the queue in a non-blocking implementation (Does it?)
SLIDE 22
Chalmers University of Technology
22
Our Approach: Sketch
- Let the head and tail pointers lag
behind the actual head and tail
- Use computation to calculate the
actual head and tail
- Trade-off between the computation
time for finding the actual head/tail and the synchronization time for keeping head/tail lag not to lag much behind.
SLIDE 23
Chalmers University of Technology
23
Results on SUN Enterprise 10000 with Full Contention
SLIDE 24
Chalmers University of Technology
24
Results on SUN Enterprise 10000 with Full Contention
SLIDE 25
Chalmers University of Technology
25
Results on SGI Origin 2000 with Full Contention
SLIDE 26
Chalmers University of Technology
26
Conclusion
- A new non-blocking concurrent FIFO queue
queue algorithm is present.
- A simple mechanism for easing ABA
problem is proposed.
- A mechanism to lower the contention of
- f non-blocking operations is introduced