|
| 1 | +Disruptor++ |
| 2 | +=========== |
| 3 | + |
| 4 | +Disruptor++ is a C++11 header-only implementation of the 'disruptor' data structure used |
| 5 | +to communicate between threads in a high-performance producer/consumer arrangement. |
| 6 | + |
| 7 | +See the LMAX [technical paper](http://disruptor.googlecode.com/files/Disruptor-1.0.pdf) |
| 8 | +for a description of the theory behind the disruptor. |
| 9 | + |
| 10 | +See the [Java LMAX Disruptor project](http://lmax-exchange.github.io/disruptor/) for |
| 11 | +more resources relating to the disruptor. |
| 12 | + |
| 13 | +Description |
| 14 | +----------- |
| 15 | + |
| 16 | +A disruptor data structure is essentially a ring buffer that uses different cursors |
| 17 | +to keep track of where consumers or producers have processed up to. |
| 18 | + |
| 19 | +A disruptor can be used with either single producer thread or multiple producer |
| 20 | +threads. Multiple producer threads are able to publish items out-of-sequence |
| 21 | +so that producer threads do not hold up other producer threads unless the |
| 22 | +entire ring buffer fills up while waiting for a slow producer to publish its item. |
| 23 | +Consumers, however, always process items in the sequence they were enqueued. |
| 24 | + |
| 25 | +The disruptor data structure supports batched operations with producers able to |
| 26 | +enqueue multiple items with a single synchronisation operation and consumers |
| 27 | +that fall behind are able to catch up by consuming batches of consecutive |
| 28 | +items with a single synchronisation operation. |
| 29 | + |
| 30 | +When used with a single producer thread with the spin wait-strategy the disruptor |
| 31 | +uses only atomic reads/writes of integers and acquire/release memory barriers for |
| 32 | +synchronisation. It does not use CAS operations or locks that require kernel |
| 33 | +arbitration. |
| 34 | + |
| 35 | +This implementation of the disruptor data-structure is defined fully in header files. |
| 36 | +There are no libraries to link against and many functions should be able to be inlined. |
| 37 | +Also, this implementation does not make use of abstract interfaces or virtual function |
| 38 | +calls, which can inhibit inlining and incur additional runtime overhead, but instead |
| 39 | +prefers to use templates for compile-time polymorphism. |
| 40 | + |
| 41 | +Performance |
| 42 | +----------- |
| 43 | + |
| 44 | +TODO: Put some performance results in here. |
| 45 | + |
| 46 | +Synopsis |
| 47 | +-------- |
| 48 | + |
| 49 | +A single producer/single consumer use of a disruptor for communication. |
| 50 | + |
| 51 | +``` |
| 52 | +#include <disruptorplus/ring_buffer.hpp> |
| 53 | +#include <disruptorplus/single_threaded_claim_strategy.hpp> |
| 54 | +#include <disruptorplus/spin_wait_strategy.hpp> |
| 55 | +#include <disruptorplus/sequence_barrier.hpp> |
| 56 | +
|
| 57 | +#include <thread> |
| 58 | +
|
| 59 | +using namespace disruptorplus; |
| 60 | +
|
| 61 | +struct Event |
| 62 | +{ |
| 63 | + uint32_t data; |
| 64 | +}; |
| 65 | +
|
| 66 | +int main() |
| 67 | +{ |
| 68 | + const size_t bufferSize = 1024; // Must be power-of-two |
| 69 | + |
| 70 | + ring_buffer<Event> buffer(bufferSize); |
| 71 | + |
| 72 | + spin_wait_strategy waitStrategy; |
| 73 | + single_threaded_claim_strategy<spin_wait_strategy> claimStrategy(bufferSize, waitStrategy); |
| 74 | + sequence_barrier<spin_wait_strategy> consumed(waitStrategy); |
| 75 | + claimStrategy.add_claim_barrier(consumed); |
| 76 | + |
| 77 | + std::thread consumer([&]() |
| 78 | + { |
| 79 | + uint64_t sum = 0; |
| 80 | + sequence_t nextToRead = 0; |
| 81 | + bool done = false; |
| 82 | + while (!done) |
| 83 | + { |
| 84 | + // Wait until more items available |
| 85 | + sequence_t available = claimStrategy.wait_until_published(nextToRead); |
| 86 | + |
| 87 | + // Process all available items in a batch |
| 88 | + do |
| 89 | + { |
| 90 | + auto& event = buffer[nextToRead]; |
| 91 | + sum += event.data; |
| 92 | + if (event.data == 0) |
| 93 | + { |
| 94 | + done = true; |
| 95 | + } |
| 96 | + } while (nextToRead++ != available); |
| 97 | + |
| 98 | + // Notify producer we've finished consuming some items |
| 99 | + consumed.publish(available); |
| 100 | + } |
| 101 | + std::cout << "sum is " << sum << std::endl; |
| 102 | + }); |
| 103 | + |
| 104 | + std::thread producer([&]() |
| 105 | + { |
| 106 | + for (uint32_t i = 1; i <= 1000000; ++i) |
| 107 | + { |
| 108 | + // Claim a slot in the ring buffer, waits if buffer is full |
| 109 | + sequence_t seq = claimStrategy.claim_one(); |
| 110 | + |
| 111 | + // Write to the slot in the ring buffer |
| 112 | + buffer[seq].data = i; |
| 113 | + |
| 114 | + // Publish the event to the consumer |
| 115 | + claimStrategy.publish(seq); |
| 116 | + } |
| 117 | + |
| 118 | + // Publish the terminating event. |
| 119 | + sequence_t seq = claimStrategy.claim_one(): |
| 120 | + buffer[seq].data = 0; |
| 121 | + claimStrategy.publish(seq); |
| 122 | + }); |
| 123 | + |
| 124 | + consumer.join(); |
| 125 | + producer.join(); |
| 126 | + |
| 127 | + return 0; |
| 128 | +} |
| 129 | +``` |
| 130 | + |
| 131 | +License |
| 132 | +------- |
| 133 | + |
| 134 | +The Discruptor++ library is available under the MIT open source license. |
| 135 | +See LICENSE.txt for details. |
| 136 | + |
| 137 | +Building |
| 138 | +-------- |
| 139 | + |
| 140 | +As this library is a header-only library there is no library component |
| 141 | +that needs to be built separately. Simply add the include/ directory |
| 142 | +to your include path and include the appropriate headers. |
| 143 | + |
| 144 | +If you want to build the samples/tests then you can use the |
| 145 | +[Cake](https://github.com/lewissbaker/cake) build system to build |
| 146 | +this code. |
| 147 | + |
| 148 | +Once you have installed cake you can simply run 'cake' in the root |
| 149 | +directory to build everything. |
0 commit comments