Circular lock-free buffer

I've made a particular study of lock-free data structures in the last couple of years. I've read most of the papers in the field (there's only about fourty or so - although only about ten or fifteen are any real use :-)

AFAIK, a lock-free circular buffer has not been invented. The problem will be dealing with the complex condition where a reader overtakes a writer or vis-versa.

If you have not spent at least six months studying lock-free data structures, do not attempt to write one yourself. You will get it wrong and it may not be obvious to you that errors exist, until your code fails, after deployment, on new platforms.

I believe however there is a solution to your requirement.

You should pair a lock-free queue with a lock-free free-list.

The free-list will give you pre-allocation and so obviate the (fiscally expensive) requirement for a lock-free allocator; when the free-list is empty, you replicate the behaviour of a circular buffer by instantly dequeuing an element from the queue and using that instead.

(Of course, in a lock-based circular buffer, once the lock is obtained, obtaining an element is very quick - basically just a pointer dereference - but you won't get that in any lock-free algorithm; they often have to go well out of their way to do things; the overhead of failing a free-list pop followed by a dequeue is on a par with the amount of work any lock-free algorithm will need to be doing).

Michael and Scott developed a really good lock-free queue back in 1996. A link below will give you enough details to track down the PDF of their paper; Michael and Scott, FIFO

A lock-free free-list is the simplest lock-free algorithm and in fact I don't think I've seen an actual paper for it.

The term of art for what you want is a lock-free queue. There's an excellent set of notes with links to code and papers by Ross Bencina. The guy whose work I trust the most is Maurice Herlihy (for Americans, he pronounces his first name like "Morris").

The requirement that producers or consumers block if the buffer is empty or full suggests that you should use a normal locking data structure, with semaphores or condition variables to make the producers and consumers block until data is available. Lock-free code generally doesn't block on such conditions - it spins or abandons operations that can't be done instead of blocking using the OS. (If you can afford to wait until another thread produces or consumes data, then why is waiting on a lock for another thread to finish updating the data structure any worse?)

On (x86/x64) Linux, intra-thread synchronization using mutexes is reasonably cheap if there is no contention. Concentrate on minimizing the time that the producers and consumers need to hold onto their locks. Given that you've said that you only care about the last N recorded data points, I think a circular buffer would be do this reasonably well. However, I don't really understand how this fits in with the blocking requirement and the idea of consumers actually consuming (removing) the data they read. (Do you want consumers to only look at the last N data points, and not remove them? Do you want producers to not care if consumers can't keep up, and just overwrite old data?)

Also, as Zan Lynx commented, you can aggregate/buffer up your data into bigger chunks when you've got lots of it coming in. You could buffer up a fixed number of points, or all the data received within a certain amount of time. This means that there will be fewer synchronization operations. It does introduce latency, though, but if you're not using real-time Linux, then you'll have to deal with that to an extent anyway.

There is a pretty good series of articles about this on DDJ. As a sign of how difficult this stuff can be, it's a correction on an earlier article that got it wrong. Make sure you understand the mistakes before you roll your own )-;

One useful technique to reduce contention is to hash the items into multiple queues and have each consumer dedicated to a "topic".

For most-recent number of items your consumers are interested in - you don't want to lock the whole queue and iterate over it to find an item to override - just publish items in N-tuples, i.e. all N recent items. Bonus points for implementation where producer would block on the full queue (when consumers can't keep up) with a timeout, updating its local tuple cache - that way you don't put back-pressure on the data source.

Sutter's queue is sub-optimal and he knows it. The Art of Multicore programming is a great reference but don't trust the Java guys on memory models, period. Ross's links will get you no definite answer because they had their libraries in such problems and so on.

Doing lock-free programming is asking for trouble, unless you want to spend a lot of time on something that you are clearly over-engineering before solving the problem (judging by the description of it, it is a common madness of 'looking for perfection' in cache coherency). It takes years and leads to not solving the problems first and optimising later, a common disease.

I would agree with this article and recommend against using lock-free data structures. A relatively recent paper on lock-free fifo queues is this, search for further papers by the same author(s); there's also a PhD thesis on Chalmers regarding lock-free data structures (I lost the link). However, you did not say how large your elements are -- lock-free data structures work efficiently only with word-sized items, so you'll have to dynamically allocate your elements if they're larger than a machine word (32 or 64 bits). If you dynamically allocate elements, you shift the (supposed, since you haven't profiled your program and you're basically doing premature optimization) bottleneck to memory allocator, so you need a lock-free memory allocator, e.g., Streamflow, and integrate it with your application.

Sometime ago, I've found a nice solution to this problem. I believe that it the smallest found so far.

The repository has a example of how use it to create N threads (readers and writers) and make then share a single seat.

I made some benchmarks, on the test example and got the following results (in million ops/sec) :

By buffer size

By number of threads

Notice how the number of threads do not change the throughput.

I think this is the ultimate solution to this problem. It works and is incredible fast and simple. Even with hundreds of threads and a queue of a single position. It can be used as a pipeline beween threads, allocating space inside the queue.

Can you break it?

Just for completeness: there's well tested lock-free circular buffer in OtlContainers, but it is written in Delphi (TOmniBaseBoundedQueue is circular buffer and TOmniBaseBoundedStack is bounded stack). There's also an unbounded queue in the same unit (TOmniBaseQueue). The unbounded queue is described in Dynamic lock-free queue – doing it right. The initial implementation of the bounded queue (circular buffer) was described in A lock-free queue, finally! but the code was updated since then.

Check out Disruptor (How to use it) which is a ring-buffer that multiple threads can subscribe to:

Here is how I would do it:

• map the queue into an array
• keep state with a next read and next next write indexes
• keep an empty full bit vector around

Insertion consists of using a CAS with an increment and roll over on the next write. Once you have a a slot, add your value and then set the empty/full bit that matches it.

Removals require a check of the bit before to test on underflows but other than that, are the same as for the write but using read index and clearing the empty/full bit.

Be warned,

1. I'm no expert in these things
2. atomic ASM ops seem to be very slow when I've used them so if you end up with more than a few of them, you might be faster to use locks embedded inside the insert/remove functions. The theory is that a single atomic op to grab the lock followed by (very) few non atomic ASM ops might be faster than the same thing done by several atomic ops. But to make this work would require manual or automatic inlineing so it's all one short block of ASM.

You may try lfqueue

It is simple to use, it is circular design lock free

int *ret;

lfqueue_t results;

lfqueue_init(&results);

/** Wrap This scope in multithread testing **/
int_data = (int*) malloc(sizeof(int));
assert(int_data != NULL);
*int_data = i++;
/*Enqueue*/
while (lfqueue_enq(&results, int_data) != 1) ;

/*Dequeue*/
while ( (ret = lfqueue_deq(&results)) == NULL);

// printf("%d\n", *(int*) ret );
free(ret);
/** End **/

lfqueue_clear(&results);

There are situations that you don't need locking to prevent race condition, especially when you have only one producer and consumer.

Consider this paragraph from LDD3:

When carefully implemented, a circular buffer requires no locking in the absence of multiple producers or consumers. The producer is the only thread that is allowed to modify the write index and the array location it points to. As long as the writer stores a new value into the buffer before updating the write index, the reader will always see a consistent view. The reader, in turn, is the only thread that can access the read index and the value it points to. With a bit of care to ensure that the two pointers do not overrun each other, the producer and the consumer can access the buffer concurrently with no race conditions.

If you take as a prerequisite that the buffer will never become full, consider using this lock-free algorithm:

capacity must be a power of 2
buffer = new T[capacity] ~ on different cache line
mask = capacity - 1
write_index ~ on different cache line
read_index ~ on different cache line

enqueue:
    write_i = write_index.fetch_add(1) & mask
    buffer[write_i] = element ~ release store

dequeue:
    read_i = read_index.fetch_add(1) & mask
    element
    while ((element = buffer[read_i] ~ acquire load) == NULL) {
        spin loop
    }
    buffer[read_i] = NULL ~ relaxed store
    return element