Skip to content

Latest commit

 

History

History
320 lines (222 loc) · 31.5 KB

README.md

File metadata and controls

320 lines (222 loc) · 31.5 KB

check-build validate codecov status

The Synch Framework

Summary

This is an open-source framework for concurrent data-structures and benchmarks. The provided framework contains a substantial set of concurrent data-structures such as queues, stacks, combining-objects, hash-tables, locks, etc. This framework also provides a user-friendly runtime for developing and benchmarking concurrent data-structures. Among other features, this runtime provides functionality for creating threads easily (both POSIX and user-level threads), tools for measuring performance, etc. The provided concurrent data-structures and the runtime are highly optimized for contemporary NUMA multiprocessors such as AMD Epyc and Intel Xeon.

The current version of this code is optimized for x86_64 machine architecture, but the code is also successfully tested in other machine architectures, such as ARM-V8 and RISC-V. Some of the benchmarks perform much better in architectures that natively support Fetch&Add instructions (e.g., x86_64, etc.).

Collection

The Synch framework provides a large set of highly efficient concurrent data-structures, such as combining-objects, concurrent queues and stacks, concurrent hash-tables and locks. The cornerstone of the Synch framework are the combining objects. A Combining object is a concurrent object/data-structure that is able to simulate any other concurrent object, e.g. stacks, queues, atomic counters, barriers, etc. The Synch framework provides the PSim wait-free combining object [2,10], the blocking combining objects CC-Synch, DSM-Synch and H-Synch [1], and the blocking combining object based on the technique presented in [4]. Moreover, the Synch framework provides the Osci blocking, combining technique [3] that achieves good performance using user-level threads. Since v3.1.0, the Synch framework offers a new high performant implementation of flat-combining synchronization technique [14]. This novel version is implemented from the scratch and is not just an optimized version of the original code provided in [15].

In terms of concurrent queues, the Synch framework provides the SimQueue [2,10] wait-free queue implementation that is based on the PSim combining object, the CC-Queue, DSM-Queue and H-Queue [1] blocking queue implementations based on the CC-Synch, DSM-Synch and H-Synch combining objects. A blocking queue implementation based on the CLH locks [5,6] and the lock-free implementation presented in [7] are also provided. Since v2.4.0, the Synch framework provides the LCRQ [11,12] queue implementation. In terms of concurrent stacks, the Synch framework provides the SimStack [2,10] wait-free stack implementation that is based on the PSim combining object, the CC-Stack, DSM-Stack and H-Stack [1] blocking stack implementations based on the CC-Synch, DSM-Synch and H-Synch combining objects. Moreover, the lock-free stack implementation of [8] and the blocking implementation based on the CLH locks [5,6] are provided. The Synch framework also provides concurrent queue and stacks implementations (i.e. OsciQueue and OsciStack implementations) that achieve very high performance using user-level threads [3]. Since v3.1.0, the Synch framework provides stack and queue implementations (i.e. FC-Stack and FC-Queue) based on the implementation of flat-combining provided by the Synch framework.

Furthermore, the Synch framework provides a few scalable lock implementations, i.e. the MCS queue-lock presented in [9] and the CLH queue-lock presented in [5,6]. Finally, the Synch framework provides two example-implementations of concurrent hash-tables. More specifically, it provides a simple implementation based on CLH queue-locks [5,6] and an implementation based on the DSM-Synch [1] combining technique.

The following table presents a summary of the concurrent data-structures offered by the Synch framework.

Concurrent Object Provided Implementations
Combining Objects CC-Synch, DSM-Synch and H-Synch [1]
PSim [2,10]
Osci [3]
Oyama [4]
FC: a new implementation of flat-combining [14]
Concurrent Queues CC-Queue, DSM-Queue and H-Queue [1]
SimQueue [2,10]
OsciQueue [3]
CLH-Queue [5,6]
MS-Queue [7]
LCRQ [11,12]
FC-Queue [14]
Concurrent Stacks CC-Stack, DSM-Stack and H-Stack [1]
SimStack [2,10]
OsciStack [3]
CLH-Stack [5,6]
LF-Stack [8]
FC-Stack [14]
Locks CLH [5,6]
MCS [9]
Hash Tables CLH-Hash [5,6]
A hash-table based on DSM-Synch [1]

Requirements

  • A modern 64-bit machine. Currently, 32-bit architectures are not supported. The current version of this code is optimized for the x86_64 machine architecture, but the code is also successfully tested in other machine architectures, such as ARM-V8 and RISC-V. Some of the benchmarks perform much better in architectures that natively support Fetch&Add instructions (e.g., x86_64, etc.).
  • A recent Linux distribution. The Synch environment may also build/run in some other Unix-like systems, (i.e. BSD, etc.). In this case the result is not guaranteed, since the environment is not tested in systems other than Linux.
  • As a compiler, gcc of version 4.8 or greater is recommended, but you may also try to use icx or clang.
  • Building requires the following development packages:
    • libatomic
    • libnuma
    • libpapi in case that the SYNCH_TRACK_CPU_COUNTERS flag is enabled in libconcurrent/config.h.
  • For building the documentation (i.e. man-pages), doxygen is required.

Configuring, compiling and installing the framework

In the libconcurrent/config.h file, the user can configure some basic options for the framework, such as:

  • Enable/disable debug mode.
  • Support for Numa machines.
  • Enable performance statistics, etc.

The provided default configuration should work well in many cases. However, the default configuration may not provide the best performance. For getting the best performance, modifying the libconcurrent/config.h may be needed (see more on Performance/Optimizations Section).

In case that you want to compile the library that provides all the implemented concurrent algorithms just execute make in the root directory of the source files. This step is necessary in case that you want to run benchmarks. However, some extra make options are provided in case the user wants to compile the framework with other than system's default compiler, clean the binary files, etc. The following table provides the list with all the available make options.

Command Description
make Auto-detects the current architecture and compiles the source-code for it (this should work for most users).
make CC=cc Compiles the source-code for the current architecture using the cc compiler.
make clang Compiles the source-code using the clang compiler.
make icx Compiles the source-code using the Intel icx compiler.
make unknown Compiles the source-code for architectures other than X86_64, e.g. RISC-V, ARM, etc.
make clean Cleaning-up all the binary files.
make docs Creating the documentation (i.e. man-pages).
make install Installing the framework on the default location (i.e. /opt/Synch/).
make install DIR=dir Installing the framework on the dir/Synch/ location.
make uninstall Uninstalling the framework.

For building the documentation (i.e. man-pages), the user should execute make docs. Notice that for building the documentation the system should be equipped with doxygen documentation tool.

For installing the framework, the user should execute make install. In this case, the framework will be installed in the default location which is /opt/Synch/. Notice that in this case, the user should have write access on the /opt directory or sudo access. The make install DIR=dir command installs the framework in the dir/Synch path, while the make uninstall uninstalls the framework. For accessing the man pages, the user should manually setup the MANPATH environmental variable appropriately (e.g. export MANPATH=$MANPATH:/opt/Synch/docs/man).

Running Benchmarks

For running benchmarks use the bench.sh script file that is provided in the main directory of this source tree.

Example usage: ./bench.sh FILE.run OPTION1 VALUE1 OPTION2 VALUE2 ...

Each benchmark reports the time that needs to be completed, the average throughput of operations performed and some performance statistics if DEBUG option is enabled during framework build. The bench.sh script measures the strong scaling of the benchmark that is executed.

The following options are available:

Option Description
-t, --max_threads set the maximum number number of POSIX threads to be used in the last set of iterations of the benchmark, default is the number of system cores
-s, --step set the step (extra number of threads to be used) in each set of iterations of the benchmark, default is number of processors/8 or 1
-f, --fibers set the number of user-level threads per POSIX thread
-r, --repeat set the total number of operations executed by the benchmark, default is 1000000
-i, --iterations set the number of times that the benchmark should be executed, default is 10
-w, --workload set the amount of workload (i.e. dummy loop iterations among two consecutive operations of the benchmarked object), default is 64
-l, --list displays the list of the available benchmarks
-n, --numa_nodes set the number of numa nodes (which may differ with the actual hw numa nodes) that hierarchical algorithms should take account
-b, --backoff, --backoff_high set an upper backoff bound for lock-free and Sim-based algorithms
-bl, --backoff_low set a lower backoff bound (only for msqueuebench, lfstackbench and lfuobjectbench benchmarks)
-h, --help displays this help and exits

The framework provides the validate.sh validation/smoke script. The validate.sh script compiles the sources in DEBUG mode and runs a big set of benchmarks with various numbers of threads. After running each of the benchmarks, the script evaluates the DEBUG output and in case of success it prints PASS. In case of a failure, the script simply prints FAIL. In order to see all the available options of the validation/smoke script, execute validate.sh -h. Given that the validate.sh validation/smoke script depends on binaries that are compiled in DEBUG mode, it is not installed while using make install. The following image shows the execution and the default behavior of validate.sh.

The framework provides another simple fast smoke test: ./run_all.sh. This will quickly run all available benchmarks with default options and store the results in the results.txt file.

Performance/Optimizations

Getting the best performance from the provided benchmarks is not always an easy task. For getting the best performance, some modifications in Makefiles may be needed (compiler flags, etc.). Important parameters for the benchmarks and/or library are placed in the libconcurrent/config.h file. A useful guide to consider in order to get better performance in a modern multiprocessor follows.

  • In case that the target machine is a NUMA machine make sure SYNCH_NUMA_SUPPORT is enabled in libconcurrent/config.h. Usually, when this option is enabled, it gives much better performance in NUMA machines. However, in some older machines this option may induce performance overheads.
  • Check the performance impact of the SYNCH_COMPACT_ALLOCATION option in libconcurrent/config.h. In modern AMD multiprocessors (i.e., equipped with EPYC processors) this option gives tremendous performance boost. In contrast to AMD processors, this option introduces serious performance overheads in Intel Xeon processors. Thus, a careful experimental analysis is needed in order to show the possible benefits of this option.
  • Check if you have selected the optimal thread placement policy (see more in Thread placement policies).
  • Check the cache line size (CACHE_LINE_SIZE and S_CACHE_LINE options in includes/system.h). These options greatly affect the performance in all modern processors. Most Intel machines behave better with CACHE_LINE_SIZE equal or greater than 128, while most modern AMD machine achieve better performance with a value equal to 64. Notice that CACHE_LINE_SIZE and S_CACHE_LINE depend on the SYNCH_COMPACT_ALLOCATION option (see includes/system.h).
  • Use backoff if it is available. Many of the provided algorithms could use backoff in order to provide better performance (e.g., sim, LF-Stack, MS-Queue, SimQueue, SimStack, etc.). In this case, it is of crucial importance to use -b (and in some cases -bl arguments) in order to get the best performance.
  • Ensure that you are using a recent gcc-compatible compiler, e.g. a gcc compiler of version 7.0 or greater is highly recommended.
  • Check the performance impact of the different available compiler optimizations. In most cases, gcc's -Ofast option gives the best performance. In addition, some algorithms (i.e., sim, osci, simstack, oscistack, simqueue and osciqueue) benefit by enabling the -mavx option (in case that AVX instructions are supported by the hardware).
  • Check if system oversubscription with user-level fibers enhances the performance. Many algorithms (i.e., the Sim and Osci families of algorithms) show tremendous performance boost by using oversubscription with user-level threads [3]. In this case, use the --fibers option.

Expected performance

The expected performance of the Synch framework is discussed in the PERFORMANCE.md file.

Thread placement policies

Since v3.2.0, the Synch framework has introduced a variety of thread placement policies. Utilizing the synchSetThreadPlacementPolicy and synchGetThreadPlacementPolicy functions, users can modify the default placement policy and inquire about the current thread placement policy, respectively. These functions enable precise control over how threads are allocated across the machine's processors, enhancing performance and efficiency. The available thread placement policies are the following:

  • SYNCH_THREAD_PLACEMENT_FLAT: Threads are distributed in a round-robin fashion across all processing cores.
  • SYNCH_THREAD_PLACEMENT_NUMA_SPARSE: Optimizes thread placement for systems with Non-Uniform Memory Access (NUMA) by spreading threads sparsely across NUMA nodes, potentially improving memory bandwidth and improving cache utilization.
  • SYNCH_THREAD_PLACEMENT_NUMA_DENSE: Places threads within the smallest number of NUMA nodes before spreading them to other nodes, which can improve memory locality and may reduce contention on shared variables.
  • SYNCH_THREAD_PLACEMENT_NUMA_DENSE_SMT_PREFER: Similar to SYNCH_THREAD_PLACEMENT_NUMA_DENSE, but with a preference for utilizing Simultaneous Multithreading (SMT) capabilities within NUMA nodes to maximize processing efficiency.
  • SYNCH_THREAD_PLACEMENT_NUMA_SPARSE_SMT_PREFER: Combines the sparse distribution strategy across NUMA nodes with a preference for SMT. This policy spreads threads across NUMA nodes to avoid contention, while preferring to fill SMT slots within each core before moving to the next. It aims to strike a balance between improving memory bandwidth and leveraging SMT for higher processing efficiency and reduced contention on shared variables.
  • SYNCH_THREAD_PLACEMENT_DEFAULT: By default the thread placement policy is se to SYNCH_THREAD_PLACEMENT_DEFAULT. Currently, SYNCH_THREAD_PLACEMENT_DEFAULT is equal to SYNCH_THREAD_PLACEMENT_NUMA_SPARSE_SMT_PREFER.

Memory reclamation (stacks and queues)

The Synch framework provides a pool mechanism (see includes/pool.h) that efficiently allocates and de-allocates memory for the provided concurrent stack and queue implementations. The allocation mechanism of this pool implementation is low-overhead. All the provided stack and queue implementations use the functionality of this pool mechanism. In order to support memory reclamation in a safe manner, a concurrent object should guarantee that each memory object that is going to de-allocated should be accessed only by the thread that is going to free it. Generally, de-allocating and thus reclaiming memory is easy in many blocking objects, since there is a lock that protects the de-allocated memory object. Currently, the Synch framework supports memory reclamation for the following concurrent stack and queue implementations:

  • Concurrent Queues:
    • CC-Queue, DSM-Queue and H-Queue [1]
    • OsciQueue [3]
    • CLH-Queue [5,6]
  • Concurrent Stacks:
    • CC-Stack, DSM-Stack and H-Stack [1]
    • OsciStack [3]
    • CLH-Stack [5,6]
    • SimStack [2,10] (since v2.4.0)

Note that de-allocating and thus recycling memory in lock-free and wait-free objects is not an easy task. Since v2.4.0, SimStack supports memory reclamation using the functionality of pool.h and a technique that is similar to that presented by Blelloch and Weiin in [13]. Notice that the MS-Queue [7], LCRQ [11,12] queue implementations and the LF-Stack [8] stack implementation support memory reclamation through hazard-pointers. However, the current version of the Synch framework does not provide any implementation of hazard-pointers. In case that a user wants to use memory reclamation in these objects, a custom hazard-pointers implementation should be integrated in the environment.

By default, memory-reclamation is enabled. In case that there is need to disable memory reclamation, the SYNCH_POOL_NODE_RECYCLING_DISABLE option should be enabled in config.h.

The following table shows the memory reclamation characteristics of the provided stack and queues implementations.

Concurrent Object Provided Implementations Memory Reclamation
Concurrent Queues CC-Queue, DSM-Queue and H-Queue [1] Supported
SimQueue [2,10] Not supported
OsciQueue [3] Supported
CLH-Queue [5,6] Supported
MS-Queue [7] Hazard Pointers (not provided by Synch)
LCRQ [11,12] Hazard Pointers (not provided by Synch)
FC-Queue [14] Supported
Concurrent Stacks CC-Stack, DSM-Stack and H-Stack [1] Supported
SimStack [2,10] Supported (since v2.4.0)
OsciStack [3] Supported
CLH-Stack [5,6] Supported
LF-Stack [8] Hazard Pointers (not provided by Synch)
FC-Stack [14] Supported

Memory reclamation limitations

In the current design of the reclamation mechanism, each thread uses a single private pool for reclaiming memory. In a producer-consumer scenario where a set of threads performs only enqueue operations (or push operations in case of stacks) and all other threads perform dequeue operations (or pop operations in case of stacks), insufficient memory reclamation is performed since each memory pool is only accessible by the thread that owns it. We aim to improve this in future versions of the Synch framework.

API documentation

A complete API documentation is provided in https://nkallima.github.io/sim-universal-construction/index.html.

Code example for a simple benchmark

We now describe a very simple example-benchmark that uses the Application Programming Interface (API) of the provided runtime. This simple benchmark measures the performance of Fetch&Add instructions in multi-core machines. The purpose of this simple benchmark is to measure the performance of Fetch&Add implementations (hardware or software).

#include <stdio.h>
#include <stdint.h>

#include <primitives.h>
#include <threadtools.h>
#include <barrier.h>

#define N_THREADS 10
#define RUNS      1000000

volatile int64_t object CACHE_ALIGN;
int64_t d1 CACHE_ALIGN, d2;
SynchBarrier bar CACHE_ALIGN;

inline static void *Execute(void *Arg) {
    long i, id;

    id = synchGetThreadId();
    synchBarrierWait(&bar);
    if (id == 0) d1 = synchGetTimeMillis();

    for (i = 0; i < RUNS; i++)
        synchFAA64(&object, 1);

    synchBarrierWait(&bar);
    if (id == 0) d2 = synchGetTimeMillis();

    return NULL;
}

int main(int argc, char *argv[]) {
    object = 1;

    synchBarrierSet(&bar, N_THREADS);
    synchStartThreadsN(N_THREADS, Execute, SYNCH_DONT_USE_UTHREADS);
    synchJoinThreadsN(N_THREADS);

    printf("time: %ld (ms)\tthroughput: %.2f (millions ops/sec)\n", 
           (d2 - d1), RUNS * N_THREADS / (1000.0 * (d2 - d1)));

    return 0;
}

This example-benchmark creates N_THREADS, where each of them executes RUNS Fetch&Add operations in a shared 64-bit integer. At the end of the benchmark the throughput (i.e. Fetch&Add operations per second) is calculated. By setting various values for N_THREADS, this benchmark is able to measure strong scaling.

The synchStartThreadsN function (provided by the API defined in threadtools.h) in main, creates N_THREADS threads and each of the executes the Execute function declared in the same file. The SYNCH_DONT_USE_UTHREADS argument imposes synchStartThreadsN to create only POSIX threads; in case that the user sets the corresponding fibers argument to M > 0, then synchStartThreadsN will create N_THREADS POSIX threads and each of them will create M user-level (i.e. fiber) threads. The synchGetThreadId (provided by threadtools.h) returns the id of the running thread, while the synchJoinThreadsN function (also provided by threadtools. h) waits until all POSIX and fiber threads (if any) finish the execution of the Execute function. The Fetch&Add instruction on 64-bit integers is performed by the synchFAA64 function provided by the API of primitives.h.

The threads executing the Execute function use the SynchBarrier re-entrant barrier object for simultaneously starting to perform Fetch&Add instructions on the shared variable object. This barrier is also re-used before the end of the Execute function in order to allow thread with id = 0 to measure the amount of time that the benchmark needed for completion. The synchBarrierSet function in main initializes the SynchBarrier object. The synchBarrierSet takes as an argument a pointer to the barrier object and the number of threads N_THREADS that are going to use it. Both synchBarrierSet and synchBarrierWait are provided by the API of barrier.h.

At the end of the benchmark, main calculates and prints the average throughput of Fetch&Add operations per second achieved by the benchmark.

If you want to cite us

@article{Kallimanis2021,
  doi = {10.21105/joss.03143},
  url = {https://doi.org/10.21105/joss.03143},
  year = {2021},
  publisher = {The Open Journal},
  volume = {6},
  number = {64},
  pages = {3143},
  author = {Nikolaos D. Kallimanis},
  title = {Synch: A framework for concurrent data-structures and benchmarks},
  journal = {Journal of Open Source Software}
}

In case you use any of the provide algorithms/data-structures in your paper, you are kindly requested to also cite the original paper that describes the algorithm/data-structure additional to the Synch framework. An exhaustive list of citations for the implemented algorithms/data-structures is provided in the References section.

Releases

An extensive list of the recent releases and their features is provided at https://github.com/nkallima/sim-universal-construction/releases.

License

The Synch framework is provided under the LGPL-2.1 License.

Code of conduct

Code of conduct.

References

[1]. Panagiota Fatourou, and Nikolaos D. Kallimanis. "Revisiting the combining synchronization technique". ACM SIGPLAN Notices. Vol. 47. No. 8. ACM, PPoPP 2012.

[2]. Panagiota Fatourou, and Nikolaos D. Kallimanis. "A highly-efficient wait-free universal construction". Proceedings of the twenty-third annual ACM symposium on Parallelism in algorithms and architectures (SPAA), 2011.

[3]. Panagiota Fatourou, and Nikolaos D. Kallimanis. "Lock Oscillation: Boosting the Performance of Concurrent Data Structures" Proceedings of the 21st International Conference on Principles of Distributed Systems (Opodis), 2017.

[4]. Yoshihiro Oyama, Kenjiro Taura, and Akinori Yonezawa. "Executing parallel programs with synchronization bottlenecks efficiently". Proceedings of the International Workshop on Parallel and Distributed Computing for Symbolic and Irregular Applications. Vol. 16. 1999.

[5]. Travis S. Craig. "Building FIFO and priority-queueing spin locks from atomic swap". Technical Report TR 93-02-02, Department of Computer Science, University of Washington, February 1993.

[6]. Peter Magnusson, Anders Landin, and Erik Hagersten. "Queue locks on cache coherent multiprocessors". Parallel Processing Symposium, 1994. Proceedings., Eighth International. IEEE, 1994.

[7]. Maged M. Michael, and Michael L. Scott. "Simple, fast, and practical non-blocking and blocking concurrent queue algorithms". Proceedings of the fifteenth annual ACM symposium on Principles of distributed computing. ACM, 1996.

[8]. R. Kent Treiber. "Systems programming: Coping with parallelism". International Business Machines Incorporated, Thomas J. Watson Research Center, 1986.

[9]. John M. Mellor-Crummey, and Michael L. Scott. "Algorithms for scalable synchronization on shared-memory multiprocessors". ACM Transactions on Computer Systems (TOCS) 9.1 (1991): 21-65.

[10]. Panagiota Fatourou, and Nikolaos D. Kallimanis. "Highly-efficient wait-free synchronization". Theory of Computing Systems 55.3 (2014): 475-520.

[11]. Adam Morrison, and Yehuda Afek. "Fast concurrent queues for x86 processors". Proceedings of the 18th ACM SIGPLAN symposium on Principles and practice of parallel programming. 2013.

[12]. Adam Morrison, and Yehuda Afek. Source code for LCRQ. http://mcg.cs.tau.ac.il/projects/lcrq.

[13]. Guy E. Blelloch, and Yuanhao Wei. "Brief Announcement: Concurrent Fixed-Size Allocation and Free in Constant Time." 34th International Symposium on Distributed Computing (DISC 2020). Schloss Dagstuhl-Leibniz-Zentrum für Informatik, 2020.

[14]. Danny Hendler, Itai Incze, Nir Shavit, and Moran Tzafrir. Flat combining and the synchronization-parallelism tradeoff. In Proceedings of the twenty-second annual ACM symposium on Parallelism in algorithms and architectures (SPAA 2010), pp. 355-364.

[15]. Danny Hendler, Itai Incze, Nir Shavit, and Moran Tzafrir. Source code for flat-combing. https://github.com/mit-carbon/Flat-Combining.

Contact

For any further information, please do not hesitate to send an email at nkallima (at) isi.gr. Feedback is always valuable.