ASCYLIB + OPTIK

ASCYLIB (with OPTIK) is a concurrent data-structure library. It contains over 40 implementations of linked lists, hash tables, skip lists, binary search trees (BSTs), queues, priority queues, and stacks. ASCYLIB contains sequential, lock-based, and lock-free implementations for each data structure.

ASCYLIB works on x86, SPARC, and Tilera architectures and contains tests to evaluate the throughput, latency, latency distribution, and energy efficiency of the included data structures.

OPTIK is a new design pattern for easily implementing fast and scalable concurrent data structures. We have merged several concurrent data structures developed with OPTIK in ASCYLIB. More details can be found here: http://lpd.epfl.ch/site/optik.

• Website : http://lpd.epfl.ch/site/ascylib - http://lpd.epfl.ch/site/optik
• Authors : Vasileios Trigonakis [email protected], Tudor David [email protected]
• Related Publications:
  • Optimistic Concurrency with OPTIK,
    Rachid Guerraoui, Vasileios Trigonakis (alphabetical order),
    PPoPP 2016
  • Asynchronized Concurrency: The Secret to Scaling Concurrent Search Data Structures,
    Tudor David, Rachid Guerraoui, Vasileios Trigonakis (alphabetical order),
    ASPLOS 2015

Algorithms

The following table contains the algorithms (and various implementations of some algorithms) included in ASCYLIB:

References

• [AKL+15] D. Alistarh, J. Kopinsky, J. Li, N. Shavit. The SprayList: A Scalable Relaxed Priority Queue. PPoPP '15.
• [BCH+10] N. G. Bronson, J. Casper, H. Chafi, and K. Olukotun. A Practical Concurrent Binary Search Tree. PPoPP '10.
• [DGT+15] T. David, R. Guerraoui, and V. Trigonakis. Asynchronized Concurrency: The Secret to Scaling Concurrent Search Data Structures. ASPLOS '15.
• [DMS+12] M. Desnoyers, P. E. McKenney, A. S. Stern, M. R. Dagenais, and J. Walpole. User-level implementations of read-copy update. PDS '12.
• [DVY+14] D. Drachsler, M. Vechev, and E. Yahav. Practical Concurrent Binary Search Trees via Logical Ordering. PPoPP '14.
• [EFR+10] F. Ellen, P. Fatourou, E. Ruppert, and F. van Breugel. Non-blocking Binary Search Trees. PODC '10.
• [F+03] K. Fraser. Practical Lock-Freedom. PhD thesis, University of Cambridge, 2004.
• [GT+16] R. Guerraoui, and V. Trigonakis. Optimistic Concurrency with OPTIK. PPoPP '16.
• [H+01] T. Harris. A Pragmatic Implementation of Non-blocking Linked Lists. DISC '01.
• [HHL+06] S. Heller, M. Herlihy, V. Luchangco, M. Moir, W. N. Scherer, and N. Shavit. A Lazy Concurrent List-Based Set Algorithm. OPODIS '05.
• [HS+12] M. Herlihy and N. Shavit. The Art of Multiprocessor Programming, Revised First Edition. 2012.
• [HLL+07] M. Herlihy, Y. Lev, V. Luchangco, and N. Shavit. A Simple Optimistic Skiplist Algorithm. SIROCCO '07.
• [HLS+11] M. Herlihy, Y. Lev, and N. Shavit. Concurrent lock-free skiplist with wait-free contains operator, May 3 2011. US Patent 7,937,378.
• [HJ+12] S. V. Howley and J. Jones. A non-blocking internal binary search tree. SPAA '12.
• [INTEL+06] Intel. Intel Thread Building Blocks. https://www.threadingbuildingblocks.org.
• [L+03] D. Lea. Overview of Package util.concurrent Release 1.3.4. http://gee.cs.oswego.edu/dl/classes/EDU/oswego/cs/dl/util/concurrent/intro.html, 2003.
• [LS+00] I. Lotan and N. Shavit. Skiplist-based concurrent priority queues. IPDPS '00.
• [M+02] M. M. Michael. High Performance Dynamic Lock-free Hash tables and List-based Sets. SPAA '02.
• [MS+96] M. M. Michael and M. L. Scott. Simple, Fast, and Practical Non-blocking and Blocking Concurrent Queue Algorithms. PODC '96.
• [NM+14] A. Natarajan and N. Mittal. Fast Concurrent Lock-free Binary Search Trees. PPoPP '14.
• [ORACLE+04] Oracle. Java CopyOnWriteArrayList. http://docs.oracle.com/javase/7/docs/api/java/util/concurrent/CopyOnWriteArrayList.html.
• [P+90] W. Pugh. Concurrent Maintenance of Skip Lists. Technical report, 1990.
• [T+86] R. Treiber. Systems Programming: Coping with Parallelism. Technical report, 1986.

New Algorithms

BST-TK is a new lock-based BST, introduced in ASCYLIB. Additionally, CLHT is a new hash hash table, introduced in ASCYLIB. We provide lock-free and lock-based variants of CLHT as a separate repository (https://github.com/LPD-EPFL/CLHT). Details of the algorithms and proofs/sketches of correctness can be found in the following technical report: https://infoscience.epfl.ch/record/203822

We have developed the following algorithms using OPTIK:

1. A simple array map (in src/hashtable-map_optik).
   We use this map in a hash table (in src/hashtable-optik0);
2. An optimistic global-lock-based linked list (in src/linkedlist-optik_gl).
   We use this list in a hash table (in src/hashtable-optik1);
3. A fine-grained linked list (in src/linkedlist-optik).
   We use this list in a hash table (in src/hashtable-optik0);
4. A skip list algorithm (in src/skiplist-optik1).
   We also provide a variant of the same algorithm (in src/skiplist-optik).

Additionally, we have optimized existing algorithms using OPTIK:

1. Java's ConcurrentHashMap algorithm (insrc/hashtable-java_optik);
2. Herlihy's optimistic skip list (in src/skiplist-optik2);
3. The classic Michael-Scott queues: * lock-based push, pop optimized with optik_lock_version_backoff (in src/queue-optik0) * lock-based push, pop optimized with optik_trylock_version (in src/queue-optik1) * lock-free push, pop optimized with optik_trylock_version (in src/queue-optik2)

Finally, we have introduced two optimization techniques inspired by OPTIK:

1. Node caching for optimizing lists (in src/linkedlist-optik_cache);
2. Victim queue for optimizing push in queues (in src/queue-optik3).

Compilation

ASCYLIB requires the ssmem memory allocator (https://github.com/LPD-EPFL/ssmem). We have already compiled and included ssmem in external/lib for x86_64, SPARC, and the Tilera architectures. Still, if you would like to create your own build of ssmem, take the following steps. Clone ssmem, do make libssmem.a and then copy libssmem.a in ASCYLIB/external/lib and smmem.h in ASCYLIB/external/include.

Additionally, the sspfd profiler library is required (https://github.com/trigonak/sspfd). We have already compiled and included sspfd in external/lib for x86_64, SPARC, and the Tilera architectures. Still, if you would like to create your own build of sspfd, take the following steps. Clone sspfd, do make and then copy libsspfd.a in ASCYLIB/external/lib and sspfd.h in ASCYLIB/external/include.

Finally, to measure power on new Intel processors (e.g., Intel Ivy Bridge), the raplread library is required (https://github.com/LPD-EPFL/raplread). We have already compiled and included raplread in external/lib. Still, if you would like to create your own build of raplread, take the following steps. Clone raplread, do make and then copy libraplread.a in ASCYLIB/external/lib and sspfd.h in ASCYLIB/external/include.

To build all data structures, you can execute make all. This target builds all lock-free, lock-based, and sequential data structures.

The last two structures, RCU and TBB, are based on external libraries. The RCU-based hash table requires an installation of the URCU library (http://urcu.so/). You need to set the URCU_PATH in common/Makefile.common to point to the folder of your local URCU installation, or alternatively, you can install URCU globally. The TBB-based hash table requires an installation of Intel's TBB library (https://www.threadingbuildingblocks.org/). You need to set the TBB_LIBS and the TBB_INCLUDES variables in common/Makefile.common, or alternatively, you can install TBB globally.

To build all data structures except from those two, you can issue make.

ASCYLIB includes a default configuration that uses gcc and tries to infer the number of cores and the frequency of the target/build platform. If this configuration is incorrect, you can always create a manual configurations in common/Makefile.common and include/utils.h (look in these files for examples). If you do not care about pinning threads to cores, these settings do not matter. You can compile with make SET_CPU=0 ... to disable thread pinning.

ASCYLIB accepts various compilation parameters. Please refer to the COMPILE file.

Tests

Building ASCYLIB generate per-data-structure benchmarks in the bin directory. Issue ./bin/executable -h for the parameters each of those accepts.

Depending on the compilation flags, these benchmarks can be set to measure throughtput, latency, and/or power-consumption statistics.

Scripts

ASCYLIB includes tons of usefull scripts (in the scripts folders). Some particularly useful ones are:

• scalability.sh and scalability_rep.h: run the given list of executable on the given (list of) number of threads, with the given parameters, and report throughput and scalability over single-threaded execution.
• scripts in apslos/ directory: they were used to create the plots for the ASPLOS '15 paper. In particular, apslos/run_scy.sh accepts configuration files (see asplos/config) so it can be configured to execute almost any per-data-structure scenario.
• scripts in ppopp/ directory: they were used to create the plots for the PPoPP '16 paper. In particular, ppopp/run_and_plot.sh can run and plot graphs for all the tests in the paper.

Thanks

Some of the initial implementations used in ASCYLIB were taken from Synchrobench (https://github.com/gramoli/synchrobench - V. Gramoli. More than You Ever Wanted to Know about Synchronization. PPoPP 2015.).