Tuesday, 07 September 2010
There have repeatedly been posts on comp.programming.threads asking for a definition of these terms. To write good multithreaded code you really need to understand what these mean, and how they affect the behaviour and performance of algorithms with these properties. I thought it would be therefore be useful to provide some definitions.
Definition of Blocking
A function is said to be blocking if it calls an operating system function that waits for an event to occur or a time period to elapse. Whilst a blocking call is waiting the operating system can often remove that thread from the scheduler, so it takes no CPU time until the event has occurred or the time has elapsed. Once the event has occurred then the thread is placed back in the scheduler and can run when allocated a time slice. A thread that is running a blocking call is said to be blocked.
Mutex lock functions such
are blocking, as are wait functions such
blocking functions are not limited to synchronization facilities:
the most common blocking functions are I/O facilities such
are also often blocking if the delay period is long enough.
Definition of Non-blocking
Non-blocking functions are just those that aren't blocking. Non-blocking data structures are those on which all operations are non-blocking. All lock-free data structures are inherently non-blocking.
Spin-locks are an example of non-blocking synchronization: if one thread has a lock then waiting threads are not suspended, but must instead loop until the thread that holds the lock has released it. Spin locks and other algorithms with busy-wait loops are not lock-free, because if one thread (the one holding the lock) is suspended then no thread can make progress.
Defintion of lock-free
A lock-free data structure is one that doesn't use any
mutex locks. The implication is that multiple threads can access the
data structure concurrently without race conditions or data
corruption, even though there are no locks — people would give
you funny looks if you suggested that
std::list was a
lock-free data structure, even though it is unlikely that there are
any locks used in the implementation.
Just because more than one thread can safely access a lock-free data structure concurrently doesn't mean that there are no restrictions on such accesses. For example, a lock-free queue might allow one thread to add values to the back whilst another removes them from the front, whereas multiple threads adding new values concurrently would potentially corrupt the data structure. The data structure description will identify which combinations of operations can safely be called concurrently.
For a data structure to qualify as lock-free, if any thread performing an operation on the data structure is suspended at any point during that operation then the other threads accessing the data structure must still be able to complete their tasks. This is the fundamental restriction which distinguishes it from non-blocking data structures that use spin-locks or other busy-wait mechanisms.
Just because a data structure is lock-free it doesn't mean that threads don't have to wait for each other. If an operation takes more than one step then a thread may be pre-empted by the OS part-way through an operation. When it resumes the state may have changed, and the thread may have to restart the operation.
In some cases, a the partially-completed operation would prevent other threads performing their desired operations on the data structure until the operation is complete. In order for the algorithm to be lock-free, these threads must then either abort or complete the partially-completed operation of the suspended thread. When the suspended thread is woken by the scheduler it can then either retry or accept the completion of its operation as appropriate. In lock-free algorithms, a thread may find that it has to retry its operation an unbounded number of times when there is high contention.
If you use a lock-free data structure where multiple threads modify the same pieces of data and thus cause each other to retry then high rates of access from multiple threads can seriously cripple the performance, as the threads hinder each other's progress. This is why wait-free data structures are so important: they don't suffer from the same set-backs.
Definition of wait-free
A wait-free data structure is a lock-free data structure with the additional property that every thread accessing the data structure can make complete its operation within a bounded number of steps, regardless of the behaviour of other threads. Algorithms that can involve an unbounded number of retries due to clashes with other threads are thus not wait-free.
This property means that high-priority threads accessing the data structure never have to wait for low-priority threads to complete their operations on the data structure, and every thread will always be able to make progress when it is scheduled to run by the OS. For real-time or semi-real-time systems this can be an essential property, as the indefinite wait-periods of blocking or non-wait-free lock-free data structures do not allow their use within time-limited operations.
The downside of wait-free data structures is that they are more complex than their non-wait-free counterparts. This imposes an overhead on each operation, potentially making the average time taken to perform an operation considerably longer than the same operation on an equivalent non-wait-free data structure.
When choosing a data structure for a given task you need to think about the costs and benefits of each of the options.
A lock-based data structure is probably the easiest to use, reason about and write, but has the potential for limited concurrency. They may also be the fastest in low-load scenarios.
A lock-free (but not wait-free) data structure has the potential to allow more concurrent accesses, but with the possibility of busy-waits under high loads. Lock-free data structures are considerably harder to write, and the additional concurrency can make reasoning about the program behaviour harder. They may be faster than lock-based data structures, but not necessarily.
Finally, a wait-free data structure has the maximum potential for true concurrent access, without the possibility of busy waits. However, these are very much harder to write than other lock-free data structures, and typically impose an additional performance cost on every access.
Design and Content Copyright © 2005-2017 Just Software Solutions Ltd. All rights reserved.