ASCYLIB (with OPTIK) is a concurrent data-structure library. The original C version (found at https://github.com/LPD-EPFL/ASCYLIB) contains over 40 implementations of linked lists, hash tables, skip lists, binary search trees (BSTs), queues, priority queues, and stacks, containing sequential, lock-based, and lock-free implementations for each data structure.
This C++ port, in its current state, contains 27 of those concurrent data structures.
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 vasileios.trigonakis@epfl.ch, Tudor David tudor.david@epfl.ch
- 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
- Optimistic Concurrency with OPTIK,
The following table contains the algorithms (and various implementations of some algorithms) included in ASCYLIB (respecting the original ASCYLIB numbering):
- [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.
- [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,
- [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.
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-Cpp/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-Cpp/external/lib
and sspfd.h
in ASCYLIB-Cpp/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-Cpp/external/lib
and sspfd.h
in ASCYLIB-Cpp/external/include
.
To build all data structures, you can execute make
.
This target builds all lock-free, lock-based, and sequential data structures.
ASCYLIB includes a default configuration that uses g++
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.
To add an algorithm, preferably implement (inherit from) one of the interfaces/abstract classes relevant to the data structure you are implementing (src/search.h
or src/stack_queue.h
). We also recommend you add it to:
- The relevant test file (
src/test_search.cc
orsrc/test_stackqueue.cc
). This means adding the algorithm to:
- The head of the file, including the relevant source header (algorithm.h)
- The
algorithms
enum near the head of the file. - The
parse_algorithm
function starting around line 123, which will parse the algorithm quick name from the command line option (-a
). - The list of algorithms in the help description, around line 550.
- The actual initialization of the appropriate data structure, starting around line 655.
-
The Makefile at the root of the project, either under
SEARCH_ALGORITHMS
orSTACKQUEUE_ALGORITHMS
. Add the name of the.h
file without the extension, as it will be added later. This ensures that a modification of your file causes recompilation of the test file. -
Don't forget to include any aux files (extra classes, or other types of data nodes) you may create to this file as well, to also ensure that the test files are recompiled if they change.
This version of ASCYLIB, unlike the C implementation, has 2 single test files.
One is for search data structures, and one is for stack/queue data structures. You can find & execute either one in the bin
directory.
Run ./bin/test_search -h
or ./bin/test_stackqueue -h
for help on the parameters (selecting the algorithm, workload, etc.)
ASCYLIB includes tons of usefull scripts (in the scripts
folders). Some particularly useful ones are:
scalability.sh
andscalability_rep.sh
: 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 (seeasplos/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.
Bear in mind that these scripts were written for the C version. Some of them have been adapted (scalability_rep_simple.sh
and also several of the run_**.sh
files inside the ppopp
directory. This should be helpful to illustrate how the change is made to run the C++ test files instead of the C ones.
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.).