Fast Wait Free Queue

This is a benchmark framework for evaluating the performance of concurrent queues. Currently, it contains four concurrent queue implementations. They are:

• A fast wait-free queue wfqueue,
• Morrison and Afek's lcrq,
• Fatourou and Kallimanis's ccqueue, and
• Michael and Scott's msqueue

The benchmark framework also includes a synthetic queue benchmark, faa, which emulates both an enqueue and a dequeue with a fetch-and-add primitive to test the performance of fetch-and-add on a system.

The framework currently contains one benchmark, pairwise, in which all threads repeatedly execute pairs of enqueue and dequeue operations. Between two operations, pairwise uses a delay routine that adds an arbitrary delay (between 50~150ns) to avoid artificial long run scenarios, where a cache line is held by one thread for a long time.

Requirements

• GCC 4.1.0 or later (Recommend GCC 4.7.3 or later): current implementations uses GCC __atomic or __sync primitives for atomic memory access.
• Linux kernel 2.5.8 or later
• glibc 2.3: we use sched_setaffinity to bind threads to cores.
• atomic CAS2: lcrq requires CAS2, a 16 Byte wide compare-and-swap primitive. This is available on most recent Intel processors and IBM Power8.
• jemalloc (optional): jemalloc eliminates the bottleneck of the memory allocator. You can link with jemalloc by setting JEMALLOC_PATH environment variable to the path where your jemalloc is installed.

How to install

Download one of the released source code tarball, then execute the following commands. The filename used may be different depending on the name of the tarball you have downloaded.

$ tar zxf fast-wait-free-queue-1.0.0.tar.gz
$ cd fast-wait-free-queue-1.0.0
$ make

This should generate 6 binaries (or 5 if your system does not support CAS2, lcrq will fail to compile): wfqueue, wfqueue0, lcrq, ccqueue, msqueue, faa, and delay. These are the pairwise benchmark compiled using different queue implementations.

• wfqueue0: the same as wfqueue except that its PATIENCE is set to 0.
• delay: a synthetic benchmark used to measure the time spent in the delay routine.

How to run

You can execute a binary directly, using the number of threads as an argument. Without an argument, the execution will use all available cores on the system.

For example,

./wfqueue 8

runs wfqueue with 8 threads.

If you would like to verify the result, compile the binary with VERIFY=1 make. Then execute a binary directly will print either PASSED or error messages.

You can also use the driver script, which invokes a binary up to 10 times and measures the mean of running times, the running time of the current run, the standard deviation, margin of error (both in time and percentage) of each run. The script terminates when the margin of error is relatively small (< 0.02), or has invoked the binary 10 times.

For example,

./driver ./wfqueue 8

runs wfqueue with 8 threads up to 10 times and collect statistic results.

You can use the benchmark script, which invokes driver on all combinations of a list of binaries and a list of numbers of threads, and report the mean running time and margin of error for each combination. You can specify the list of binaries using the environment variable TESTS. You can specify the list of numbers of threads using the environment variable PROCS.

The generated output of benchmark can be used as a datafile for gnuplot. The first column of benchmark's output is the number threads. Then every two columns are the mean running time and margin of error for each queue implementation. They are in the same order as they are specified in TESTS.

For example,

TESTS=wfqueue:lcrq:faa:delay PROCS=1:2:4:8 ./benchmark

runs each of wfqueue, lcrq, faa, and delay using 1, 2, 4, and 8 threads.

Then you can plot them using,

set logscale x 2
plot "t" using 1:(20000/($2-$8)) t "wfqueue" w lines, \
     "t" using 1:(20000/($4-$8)) t "lcrq" w lines, \
     "t" using 1:(20000/($6-$8)) t "faa" w lines

How to map threads to cores

By default, the framework will map a thread with id i to the core with id i % p, where p is the number of available cores on a system; you can check each core's id in proc/cpuinfo.

To implement a custom mapping, you can add a cpumap function in cpumap.h. The signature of cpumap is

int cpumap(int id, int nprocs)

where id is the id of the current thread, nprocs is the number of threads. cpumap should return the corresponding core id for the thread. cpumap.h contains several examples of the cpumap function. You should guard the definition of the added cpumap using a conditional macro, and add the macro to CFLAGS in the makefile.

How to add a new queue implementation

We use a generic pointer void * to represent a value that can be stored in the queue. A queue should implements the queue interface, defined in queue.h.

• queue_t: the struct type of the queue,
• handle_t: a thread's handle to the queue, used to store thread local state,
• void queue_init(queue_t * q, int nprocs): initialize a queue; this will be called only once,
• void queue_register(queue_t * q, handle_t * th, int id): initialize a thread's handle; this will be called by every thread that uses the queue,
• void enqueue(queue_t * q, handle_t * th, void * val): enqueues a value,
• void * dequeue(queue_t * q, handle_t * th): dequeues a value,
• void queue_free(queue_t * q, handle_t * h): deallocate a queue and cleanup all resources associated with it,
• EMPTY: a value that will be returned if a dequeue fails. This should be a macro that is defined in the header file.

How to add a new benchmark

A benchmark should implement the benchmark interface, defined in benchmark.h, and interact with a queue using the queue interface. The benchmark interface includes:

• void init(int nprocs, int n): performs initialization of the benchmark; called only once at the beginning.
• void thread_init(int id, int nprocs): performs thread local initialization of the benchmark; called once per thread, after init but before benchmark.
• void * benchmark(int id, int nprocs): run the benchmark once, called by each thread to run the benchmark. Each call will be timed and report as one iteration. It can return a result, which will be passed to verify to verify correctness.
• int verify(int nprocs, void * results): should verify the result of each thread and return 0 on success and non-zero values on error.