Algebraic-Programming / ALP

Home of ALP/GraphBLAS and ALP/Pregel, featuring shared- and distributed-memory auto-parallelisation of linear algebraic and vertex-centric programs. Soon with more to come!
Apache License 2.0
25 stars 4 forks source link
   _____  .____   __________
  /  _  \ |    |  \______   \
 /  /_\  \|    |   |     ___/
/    |    \    |___|    |
\____|__  /_______ \____|
        \/        \/

  Copyright 2021 Huawei Technologies Co., Ltd.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

This distribution contains the C++ Algebraic Programming (ALP) framework, and provides the ALP/GraphBLAS, ALP/Pregel, and Sparse BLAS programming interfaces. Only a subset of Sparse BLAS functionality is supported, at present.

This distribution contains ALP backends that generate:

Additional backends may optionally be enabled by providing their dependences. Those backends generate:

All backends perform automatically generate vectorised programs, amongst other automatically-applied optimisations.

The ALP/GraphBLAS and ALP/Pregel interfaces are enabled for all backends, while the standard Sparse BLAS APIs only allow for the efficient support of the sequential and shared-memory parallel backends.

Minimal requirements

We first summarise the compile-time, link-time, and run-time dependences of ALP. The following are required for producing both sequential and shared-memory ALP libraries and programs, using its reference and reference_omp backends.

Compilation

To compile ALP, you need the following tools:

  1. A C++11-capable compiler such as GCC 4.8.2 or higher, with OpenMP support
  2. LibNUMA development headers
  3. POSIX threads development headers
  4. CMake version 3.13 or higher, with GNU Make (CMake's default build tool on UNIX systems) or any other supported build tool.

Linking and run-time

The ALP libraries link against the following libraries:

  1. LibNUMA: -lnuma
  2. Standard math library: -lm
  3. POSIX threads: -lpthread
  4. OpenMP: -fopenmp in the case of GCC

Optional dependences

The below summarises the dependences for optional features.

Distributed-memory auto-parallelisation

For distributed-memory parallelisation, the Lightweight Parallel Foundations (LPF) communication layer, version 1.0 or higher, is required. ALP makes use of the LPF core library and its collectives library. The LPF library has its further dependences, which are all summarised on the LPF project page:

The dependence on LPF applies to compilation, linking, and run-time. Fulfilling the dependence enables the bsp1d and hybrid ALP/GraphBLAS backends.

Code documentation

For generating the code documentations:

Code coverage

For code/test coverage, a native implementation is available using the CMake infrastructure, using gcovr and gcov/lcov.

Very quick start

Here are example steps to compile and install ALP for shared-memory machines without distributed-memory support. The last three commands show-case the compilation and execution of the sp.cpp example program.

cd <ALP/root/dir>
mkdir build
cd build
../bootstrap.sh --prefix=../install
make -j
make -j install
source ../install/bin/setenv
grbcxx ../examples/sp.cpp
grbrun ./a.out

Quick start

In more detail, the steps to follow are:

  1. Edit the include/graphblas/base/config.hpp. In particular, please ensure that config::SIMD_SIZE::bytes defined in that file is set correctly with respect to the target architecture.

  2. Create an empty directory for building ALP and move into it: mkdir build && cd build.

  3. Invoke the bootstrap.sh script located inside the ALP root directory <ALP/root/dir> to generate the build infrastructure via CMake inside the the current directory:

    <ALP/root/dir>/bootstrap.sh --prefix=</path/to/install/dir>

    • note: add --with-lpf=/path/to/lpf/install/dir if you have LPF installed and would like to use it.
  4. Issue make -j to compile the C++11 ALP library for the configured backends.

  5. (Optional) To later run all unit tests, several datasets must be made available. Please run the <ALP/root/dir>/tools/downloadDatasets.sh script for

    a. an overview of datasets required for the basic tests, as well as

    b. the option to automatically download them.

  6. (Optional) To make the ALP documentation, issue make userdocs. This generates both

    a. LaTeX in <ALP build dir>/docs/user/latex/refman.tex, and

    b. HTML in <ALP build dir>/docs/user/html/index.html.

    To build a PDF from the LaTeX sources, cd into the directory mentioned, and issue make.

  7. (Optional) Issue make -j smoketests to run a quick set of functional tests. Please scan the output for any failed tests. If you do this with LPF enabled, and LPF was configured to use an MPI engine (which is the default), and the MPI implementation used is not MPICH, then the default command lines the tests script uses are likely wrong. In this case, please edit tests/parse_env.sh by searching for the MPI implementation you used, and uncomment the lines directly below each occurrence.

  8. (Optional) Issue make -j unittests to run an exhaustive set of unit tests. Please scan the output for any failed tests. If you do this with LPF enabled, please edit tests/parse_env.sh if required as described in step 5.

  9. Issue make -j install to install ALP into the install directory configured during step 1.

  10. (Optional) Issue source </path/to/install/dir>/bin/setenv to make available the grbcxx and grbrun compiler wrapper and runner.

Congratulations, you are now ready for developing and integrating ALP algorithms! Any feedback, question, problem reports are most welcome at

albertjan.yzelman@huawei.com

Additional Contents

The remainder of this file summarises configuration options, additional build system targets, how to integrate ALP programs into applications, debugging, and contribute to ALP development. Finally, this README acknowledges contributors and lists technical papers.

Configuration

ALP employs configuration headers that contain constexpr settings that take effect every time ALP programs are compiled. Multiple object files that were compiled using ALP must all been compiled using the same configuration settings-- linking objects that have been compiled with a mixture of configurations are likely to incur undefined behaviour. The recommendation is to set a configuration before building and installing ALP, and to keep the installation directories read-only so that configurations remain static.

There exists one main configuration file that affects all ALP backends, while other configuration files only affect a specific backend or only affect specific classes of backends. The main configuration file is found in <root>/include/graphblas/base/config.hpp, which allows one to set the

  1. cache line size, in bytes, within the CACHE_LINE_SIZE class;
  2. SIMD width, in bytes, within the SIMD_SIZE class;
  3. default number of experiment repetitions during benchmarking, within the BENCHMARKING class;
  4. L1 data cache size, in bytes, within MEMORY::big_memory class;
  5. from which size onwards memory allocations will be reported, in log-2 bytes, within MEMORY::big_memory;
  6. index type used for row coordinates, as the RowIndexType typedef;
  7. index type used for column coordinates, as the ColIndexType typedef;
  8. type used for indexing nonzeroes, as the NonzeroIndexType typedef;
  9. index type used for vector coordinates, as the VectorIndexType typedef.

Other configuration values in this file are automatically inferred, are fixed non-configurable settings, or are presently not used by any ALP backend.

Reference and reference_omp backends

The file include/graphblas/reference/config.hpp contain defaults that pertain to the auto-vectorising and sequential reference backend, but also to the shared-memory auto-parallelising reference_omp backend. It allows one to set

  1. whether prefetching is enabled in PREFETCHING::enabled;
  2. the prefetch distance in PREFETCHING::distance;
  3. the default memory allocation strategy for thread-local data in IMPLEMENTATION::defaultAllocMode();
  4. same, but for shared data amongst threads in IMPLEMENTATION::sharedAllocMode();

Modifying any of the above should be done with utmost care as it typically affects the defaults across an ALP installation, and all programs compiled using it. Configuration elements not mentioned here should not be touched by users, and rather should concern ALP developers only.

OpenMP backends

The file include/graphblas/omp/config.hpp contains some basic configuration parameters that affect any OpenMP-based backend. However, the configuration file does not contain any other user-modifiable settings, but rather contains a) some utilities that OpenMP-based backends may rely on, and b) default that are derived from other settings described in the above. These settings should only be overridden with compelling and expert knowledge.

LPF backends

The file include/graphblas/bsp/config.hpp contains some basic configuration parameters that affect any LPF-based backend. It includes:

  1. an initial maximum of LPF memory slot registrations in LPF::regs();
  2. an initial maximum of LPF messages in LPF::maxh().

These defaults, if insufficient, will be automatically resized during execution. Setting these large enough will therefore chiefly prevent buffer resizes at run- time. Modifying these should normally not lead to significant performance differences.

Utilities

The file include/graphblas/utils/config.hpp details configurations of various utility functions, including:

  1. a buffer size used during reading input files, in PARSER::bsize();
  2. the block size of individual reads in PARSER::read_bsize().

These defaults are usually fine except when reading from SSDs, which would benefit of a larger read_bsize.

Others

While there are various other configuration files (find config.hpp), the above should list all user-modifiable configuration settings of interest. The remainder pertain to configurations that are automatically deduced from the aforementioned settings, or pertain to settings that describe how to safely compose backends and thus only are of interest to ALP developers.

Overview of the main Makefile targets

The following table lists the main build targets of interest:

Target Explanation
[default] builds the ALP libraries and examples, including
Sparse BLAS libraries generated by ALP
install install libraries, headers and some convenience
scripts into the path set via --prefix=<path>
unittests builds and runs all available unit tests
smoketests builds and runs all available smoke tests
perftests builds and runs all available performance tests
tests builds and runs all available unit, smoke, and
performance tests
userdocs builds HTML and LaTeX documentation corresponding
to the public ALP API
devdocs builds HTML and LaTeX code documentation for
developers of the ALP internals
docs build both the user and developer code
documentation

For more information about the testing harness, please refer to the related documentation.

For more information on how the build and test infrastructure operate, please refer to the the related documentation.

Automated performance testing

To check in-depth performance of this ALP implementation, issue make -j perftests. This will run several algorithms in several ALP configurations. This generates three main output files:

  1. <ALP/build/dir>/tests/performance/output, which summarises the whole run;

  2. <ALP/build/dir>/tests/performance/output/benchmarks, which summarises the performance of individual algorithms; and

  3. <ALP/build/dir>/tests/performance/output/scaling, which summarises operator scaling results.

To ensure that all tests run, please ensure that all related datasets are available, as also described at step 5 of the quick start.

With LPF enabled, please note the remark described at steps 3 and 7 of the quick start guide. If LPF was not configured using MPICH, please review and apply any necessary changes to tests/performance/performancetests.sh.

Integrating ALP with applications

There are several use cases in which ALP can be deployed and utilised, listed in the following. These assume that the user has installed ALP in a dedicated directory via make install.

Running ALP programs as standalone executables

Implementation

The grb::Launcher< AUTOMATIC > class abstracts a group of user processes that should collaboratively execute any single ALP program. The ALP program of interest must have the following signature:

void grb_program( const T& input_data, U& output_data )

The types T and U can be any plain-old-data (POD) type, including structs -- these can be used to broadcast input data from the master process to all user processes (input_data) -- and for data to be sent back on exit of the parallel ALP program.

The above sending-and-receiving across processes applies only to ALP implementations and backends that support or require multiple user processes; both the sequential reference and the shared-memory parallel reference_omp backends, for example, support only one user process.

In case of multiple user processes, the overhead of the broadcasting of input data is linear in the number of user processes, as well as linear in the byte- size of T which hence should be kept to a minimum. A recommended use of this mechanism is, e.g., to broadcast input data locations; any additional I/O should use the parallel I/O mechanisms that ALP exposes to the ALP program itself.

Output data is retrieved only from the user process with ID 0, even if multiple user processes exist. Some implementations or systems may require sending back the output data to a calling process, even if there is only one user process. The data movement cost incurred should hence be considered linear in the byte size of U, and, similar to the input data broadcasting, the use of parallel I/O facilities from the ALP program itself for storing large outputs is strongly advisable.

Compilation

Our backends auto-vectorise, hence please recall step 1 from the quick start guide, and make sure the include/graphblas/base/config.hpp file reflects the correct value for config::SIMD_SIZE::bytes. This value must be updated prior to the compilation and installation of ALP.

When targeting different architectures with differing SIMD widths, different ALP installations for different architectures could be maintained.

ALP programs may be compiled using the compiler wrapper grbcxx that is generated during installation. To compile high-performance code when compiling your programs using the ALP installation, the following flags are recommended:

Omitting these flags for brevity, some compilation examples follow.

When using the LPF-enabled hybrid shared- and distributed-memory ALP backends,

grbcxx -b hybrid

as the compiler command. To show all flags that the wrapper passes on, please use

grbcxx -b hybrid --show

and append your regular compilation arguments.

The hybrid backend is capable of spawning multiple ALP user processes. In contrast, compilation using

grbcxx -b reference

produces a sequential binary, while

grbcxx -b reference_omp

produces a shared-memory parallel binary.

Note that the ALP source code never requires change while switching backends.

Linking

The executable must be statically linked against an ALP library that is different depending on the selected backend. The compiler wrapper grbcxx takes care of all link-time dependencies automatically. When using the LPF-enabled BSP1D backend to ALP, for example, simply use grbcxx -b bsp1d as the compiler/linker command.

Use

grbcxx -b bsp1d --show <your regular compilation command>

to show all flags that the wrapper passes on.

Running

The resulting program has run-time dependencies that are taken care of by the LPF runner lpfrun or by the ALP runner grbrun.

We recommend using the latter:

grbrun -b hybrid -np <P> </path/to/my/program>

Here, P is the number of requested ALP user processes.

Threading

The hybrid backend employs threading in addition to distributed-memory parallelism. To employ threading to use all available hyper-threads or cores on a single node, the reference_omp backend may be selected instead.

In both cases, make sure that during execution the OMP_NUM_THREADS and OMP_PROC_BIND environment variables are set appropriately on each node that executes ALP user process(es).

Running parallel ALP programs from existing parallel contexts

This, instead of automatically spawning a requested number of user processes, assumes a number of processes already exist and that we wish those processes to jointly execute a single parallel ALP program.

Implementation

The binary that contains the ALP program to be executed must define the following global symbol with the given value:

const int LPF_MPI_AUTO_INITIALIZE = 0

A program may then again be launched via the Launcher, but in this case the MANUAL template argument should be used instead. This specialisation disallows the use of a default constructor. Instead, construction requires four arguments as follows:

grb::Launcher< MANUAL > launcher( s, P, hostname, portname )

Here, P is the total number of processes that should jointly execute a parallel ALP program, while 0 <= s < P is a unique ID of this process amongst its P-1 siblings. The types of s and P are size_t, i.e., unsigned integers.

One of these processes must be selected as a connection broker prior to forming a group of ALP user processes. The remainder P-1 processes must first connect to the chosen broker using TCP/IP connections. This choice must be made outside of ALP, prior to setting up the launcher, and materialises as the hostname and portname Launcher constructor arguments. The host and port name are strings, and must be equal across all processes.

As before, and after the successful construction of a manual launcher instance, a parallel ALP program is launched via

grb::Launcher< MANUAL >::exec( &grb_program, input, output )

in exactly the same way as described earlier, though with the input and output arguments now being passed in a one-to-one fashion:

  1. The input data is passed on from the original process to exactly one corresponding ALP user process; i.e., no broadcast occurs. The original process and the ALP user process are, from an operating system point of view, the same process. Therefore, and additionally, input no longer needs to be a plain-old-data (POD) type. Pointers, for example, are now perfectly valid to pass along, and enable sharing data between the original process and the ALP algorithm.
  2. The output data is passed from each ALP user process to the original process that called Launcher< MANUAL >::exec. To share ALP vector data, it is, for example, legal to return a grb::PinnedVector< T > as the exec output argument type. Doing so is akin to returning a pointer to output data, and does not explicitly pack nor transmit vector data.

Running

The pre-existing process must have been started using an external mechanism. This mechanism must include run-time dependence information that is normally passed by the ALP runner whenever a distributed-memory parallel backend is selected.

If the external mechanism by which the original processes are started allows it, this is most easily effected by using the standard grbcxx launcher while requesting only one process only, e.g.,

grbrun -b hybrid -n 1 </your/executable>

If the external mechanism does not allow this, then please execute e.g.

grbrun -b hybrid -n 1 --show </any/executable>

to inspect the run-time dependences and environment variables that must be made available, resp., set, as part of the external mechanism that spawns the original processes.

Integrating ALP within your coding project

Please see this article on how to add ALP and ALP/GraphBLAS as a dependence to your project.

Debugging

To debug an ALP program, please compile it using the sequential reference backend and use standard debugging tools such as valgrind and gdb. Additionally, please ensure to not pass the -DNDEBUG flag during compilation.

If bugs appear in one backend but not another, it is likely you have found a bug in the former backend. Please send a minimum working example that demonstrates the bug to the maintainers, either as an issue on or an email to:

  1. GitHub;
  2. Gitee;
  3. Albert-Jan.

Development in ALP

Your contributions to ALP would be most welcome. Merge Requests (MRs) can be contributed via Gitee and GitHub; see above for the links.

For the complete development documentation, you should start from the docs/README file and the related Development guide.

Acknowledgements

The LPF communications layer was primarily authored by Wijnand Suijlen, without whom the current ALP would not be what it is now.

The collectives library and its interface to the ALP was primarily authored by Jonathan M. Nash.

The testing infrastructure that performs smoke, unit, and performance testing of sequential, shared-memory parallel, and distributed-memory parallel backends was primarily developed by Daniel Di Nardo.

ALP and ALP/GraphBLAS have since developed significantly, primarily through efforts by researchers at the Huawei Paris and Zürich Research Centres, and the Computing Systems Laboratory in Zürich in particular. See the NOTICE file for individual contributors.

Citing ALP, ALP/GraphBLAS, and ALP/Pregel

If you use ALP in your work, please consider citing one or more of the following papers, as appropriate.

ALP and ALP/GraphBLAS

ALP/Pregel