Skip to content

Latest commit

 

History

History
623 lines (492 loc) · 22.2 KB

README.md

File metadata and controls

623 lines (492 loc) · 22.2 KB

GitHub Releases License: MIT Documentation Build Status Coverage status

MPsym - Multiprocessor System Symmetry Reduction

Contents

Introduction

MPsym is a C++/Python library that makes it possible to determine whether mappings of computational tasks to multiprocessor systems are equivalent by symmetry. This is useful when trying to find an optimal/good (with respect to runtime/energy consumption etc.) mapping or set of mappings to a given system. MPsym is able to implicitly partition the search space of all possible mappings into equivalence classes of mappings that have almost identical runtime properties due to architectural symmetries. A search algorithm can then work with representatives of these equivalence classes, effectively reducing the size of the search space ifself.

MPsym uses algorithms and data structures from computational group theory as presented in e.g. [1]. These are implemented from scratch to avoid reliance on existing computational algebra systems like GAP [2]. As a result, MPsym can be more easily integrated with other C++ and Python applications and could be released under the liberal MIT License.

Although MPsym has been developed in the context of multiprocessor systems, it is potentially applicable to other problems involving symmetries (more formally: automorphism groups) of graphs.

A Motivating Example

As an introductory example, consider an abstract Multiprocessor System-on-Chip (MPSoC) architecture consisting of four processing elements connected circularly by four bidirectional communication channels (which could represent direct wired/wireless links, shared memory etc.). We can represent this architecture by the following architecture graph:

Every vertex corresponds to a processing element and each edge corresponds to a communication channel. In general, such architecture graphs might also be vertex- and or edge-colored to reflect non-identical processing elements and communication channels.

Say we want to map two tasks T_red and T_green to this architecture. Assuming that we don't wish to map both of them to the same processing element, there are twelve distinct ways of doing so:

We refer to this set of all possible mappings for a given set of tasks as the full mapping space. MPsym is able to (implicitly) partition the full mapping space into sets of mappings equivalent by symmetry:

We refer to such sets of mappings equivalent by symmetry as orbits. MPsym collapses the full mapping space by reducing each orbit to one arbitrary mapping contained in it. If we represent mappings by tuples of processor indices, it is natural to choose the lexicographically smallest such mapping (as shown above) which we call the canonical representative.

Since explicitly enumerating all orbits for a given mapping space is often prohibitively expensive, MPsym offers functions for directly determining the canonical representative for any given mapping.

How It Works

For simple architectures, MPsym performs the following steps in order to determine canonical representatives for a number of mappings:

  1. Parse an architecture configuration file.
  2. Construct a totally colored architecture graph.
  3. Determine that architecture graph's automorphism group using nauty [3].
  4. Construct a base and strong generating set representation for this group.
  5. Find the canonical representative for a given mapping by:
    • Enumerating the orbit.
    • Enumerating the automorphism group.
    • Using local search.

The automorphism group of an architecture graph captures all of the graphs inherent symmetries in the form of permutations. For certain classes of hierarchical architectures, MPsym uses the methods presented in [4] to speed up this process, making it viable for large but highly symmetrical architectures.

Installation

This section explains how to install MPsym to your system. Note that MPsym currently only runs on Linux. It should in principle be possible to build it on non-Linux systems but this is currently either untested or simply not yet supported by the build system.

Via pip

If you only plan on calling MPsym from Python, then the easiest way to install it is via pip install pympsym. This requires Python >= 3.6 and pip >= 19.3.

From Source

If you plan to install MPsym from source you will need the following installed on your system:

  • CMake >= 3.6
  • Boost >= 1.72.0
  • Lua >= 5.2.0
  • LuaRocks

If you only plan on calling MPsym from Python simply run pip install . or pip install --user ..

If you want to call MPsym from C++, you need to directly build MPsym using CMake. Run the following from the root of this repository:

mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/usr/local
make -j $(nproc)

Afterwards, run make install to install the C++ header files and shared objects as well as the mpsym Lua rock to your system. If you do not want to install the Lua rock you can pass -DLUA_EMBED=ON to CMake to embed the mpsym Lua module into a shared object. If you do not want to use Lua architecture graph configuration files at all, you can instead pass DLUA_NO_ROCK=ON to CMake.

You can also pass -DPYTHON_BINDINGS=ON to CMake to additionally install the Python bindings without separately invoking pip.

Examples

The following brief examples showcase how to use the Python interface of MPsym. For more examples, in both Python and C++, take a look at the unit tests under test/tests.

Defining Architecture Graphs

MPsym can parse architecture graph descriptions given as Lua scripts or JSON files (the latter mostly for serialization purposes). It is also possible to construct architecture graphs programmatically.

The aforementioned Lua scripts must return a table describing an architecture graph. This table is constructed using the mpsym Lua module. This module defines several convenience functions and tables which can be used to quickly construct complex architecture graphs via the following steps:

  1. Define a set of processing elements.
  2. Connect them with communication channels.
  3. Construct an ArchGraph table.
  4. (Optional) Repeat steps 1-4 to construct additional ArchGraph tables.
  5. (Optional) Compose the constructed ArchGraph tables using ArchGraphCluster and ArchUniformSuperGraph.

Here, ArchGraphCluster represents a set of architecture graphs "isolated" from each other, such that mappings within an orbit never map the same task to different architecture graphs within the cluster. In contrast , ArchUniformSuperGraph represents a "super graph", i.e. a set of identical architecture graphs connected among each other.

As a simple example, let's consider again the architecture graph from our introductory example. We can construct it as follows:

local mpsym = require 'mpsym'

return mpsym.ArchGraph:create{
  directed = false,
  processors = {
    {0, 'P'},
    {1, 'P'},
    {2, 'P'},
    {3, 'P'}
  },
  channels = {
    {0, 1, 'C'},
    {1, 2, 'C'},
    {2, 3, 'C'},
    {3, 0, 'C'}
  }
}

Here, the integers are processor indices and P / C are arbitrary processing element and communication channel-type labels. The exact form of these labels is not important, but it is cruccial, that identical processing elements/communication channels receive identical labels.

Explicitly specifying all processing elements and communication channels can be tedious and error prone for more complex architecture graphs. While it is often possible to more succinctly construct the respective tables using Lua language features, the mpsym module also offers several convenience functions for this purpose. For instance, the architecture graph from the previous example can be more easily constructed as such:

local mpsym = require 'mpsym'

local processors = mpsym.identical_processors(4, 'P')
local channels = mpsym.grid_channels(processors, 'C')

return mpsym.ArchGraph:create{
  directed = false,
  processors = processors,
  channels = channels
}

We can parse a Lua configuration file in Python as follows:

import pympsym

ag = pympsym.ArchGraphSystem.from_lua_file('arch_graph.lua')

We can also explicitly construct architecture graphs, e.g.:

import pympsym

ag = pympsym.ArchGraph()

ag.add_processors(4, 'P')

for i in range(4):
    ag.add_channel(i, (i + 1) % 4, 'C')

Hierarchical Architecture Graphs

MPsym can determine representatives especially efficiently when working with certain hierarchical graph. For example, Kalray's MPPA3 Coolidge processor consists or sixteen identical clusters of processing elements which are fully connected internally (via shared memory) and connected in a grid fashion among each other. We can use ArchUniformSuperGraph to model this (here "proto" refers to the architecture of each cluster and "super" to the interconnections between them):

local mpsym = require 'mpsym'

local super_graph_clusters = mpsym.identical_clusters(16, 'SoC')
local super_graph_channels = mpsym.grid_channels(super_graph_clusters, 'C')

local proto_processors = mpsym.identical_processors(16, 'P')
local proto_channels = mpsym.fully_connected_channels(proto_processors, 'shared memory')

return mpsym.ArchUniformSuperGraph:create{
  super_graph = mpsym.ArchGraph:create{
    directed = false,
    clusters = super_graph_clusters,
    channels = super_graph_channels
  },
  proto = mpsym.ArchGraph:create{
    directed = false,
    processors = proto_processors,
    channels = proto_channels
  }
}

This also works in Python:

import pympsym

ag_super = pympsym.ArchGraph()
# ... build super graph

ag_proto = pympsym.ArchGraph()
# ... build proto graph

ag = pympsym.ArchUniformSuperGraph(ag_super, ag_proto)

Both ArchGraph and ArchUniformSuperGraph are subclasses of the abstract ArchGraphSystem class which defines methods for determining orbits, representatives etc. There is also ArchGraphCluster which combines an arbitrary number of different and unconnected architecture graphs.

Persisting Architecture Graphs

Once constructed, it's possible to convert ArchGraphSystem objects to and from JSON. This is especially useful because initializing an architecture graph after its construction can require a non-trivial amount of computing time which can be avoided on subsequent runs by loading an already initialized architecture graph from a JSON file.

>>> ag.to_json()
'{"automorphisms": [4,[1, 0],["(0, 1)(2, 3)", "(0, 3)", "(1, 2)"]]}'
>>> ag = pympsym.ArchGraphSystem.from_json(...)

Initializing Architecture Graphs

Before we can perform any useful operations on an ArchGraphSystem object, we need to initialize it. This is a separate step that must be performed exactly once after construction (and again if the architecture graph changes):

>>> ag.initialize()

Note that it is not necessary to call ArchGraphSystem.initialize explicitly. Operations that require it will call this method implicitly. However, since it might take a significant amount of time to complete, it can be more sensible to separate initialization and use of an architecture graph. We can also pass a timeout argument, such that an exception will be raised if initialization does not complete in the given number of seconds:

>>> ag.initialize(timeout=2.5) # raise exception after 2.5 seconds of runtime

Several other ArchGraphSystem methods also take a timeout parameter that works the same way.

Orbits and Representatives

Given an architecture graph, we can easily determine the orbit of an arbitrary mapping:

>>> list(ag.orbit((0,1)))
[(0, 1), (0, 3), (1, 0), (1, 2), (2, 1), (2, 3), (3, 0), (3, 2)]
>>> list(ag.orbit((0,2)))
[(0, 2), (1, 3), (2, 0), (3, 1)]

Orbits are constructed lazily, i.e. the orbit elements are determined incrementally while iterating through the object returned by ArchGraphSystem.orbit. The lexicographically smallest mapping in each orbit is its representative. We can also directly determine this representative (possibly much more efficiently) as follows:

>>> ag.representative((1,0))
(0, 1)

The method argument controls how the representative is determined. iterate and orbit always produce the correct representative and which one is faster depends on the given architecture graph and mapping. local_search_bfs and local_search_dfs are very fast, but the returned representative is not guaranteed to be correct (the likelihood of an incorrect result again varies with architecture graphs and mappings):

>>> ag.representative((1,0), method='auto') # default
(0, 1)
>>> ag.representative((1,0), method='iterate') # iterate through automorphism group
(0, 1)
>>> ag.representative((1,0), method='orbit') # enumerate orbit
(0, 1)
>>> ag.representative((1,0), method='local_search_bfs') # BFS local search
(0, 1)
>>> ag.representative((1,0), method='local_search_dfs') # DFS local search
(0, 1)

ArchGraphSystem.representative also takes an optional parameter of type Representatives which conveniently stores all determined representatives and causes ArchGraphSystem.representative to return a boolean flag and an integer in addition to the determined representative. The boolean flag indicates whether or not the representative has not been encountered before and the integer uniquely identifies the orbit which the representative belongs to.

>>> representatives = pympsym.Representatives
>>> ag.representative((0, 1), representatives)
((0, 1), True, 0)
>>> ag.representative((0, 2), representatives)
((0, 2), True, 1)
>>> ag.representative((0, 3), representatives)
((0, 1), False, 0)

This makes it possible to pass a number of mappings to ArchGraphSystem.representative in sequence and to immediately decide whether or not to e.g. skip a computationally expensive simulation step if the current mapping is equivalent by symmetry to a previously simulated one:

mappings = [...]

simulation_results = {}

for mapping in mappings:
    _, new, index = ag.representative(mapping)

    if new:
        simulation_results[index] = simulate(mapping)

    print('simulation results: {}'.format(simulation_results[index]))

Automorphism Groups

We can directly retrieve the automorphism group of an ArchGraphSystem object:

>>> pg = ag.automorphisms()

The returned object is of type PermGroup and acts like a set of Perm objects:

>>> pg.degree()
4
>>> len(pg)
8
>>> next(iter(pg))
(0, 1)(2, 3)
>>> '(1,2)' in pg
True
>>> '(1,3)' in pg
False

The "degree" of a permutation group is the largest element it acts on + 1, for an architecture graph's automorphism group this corresponds to the number of processing elements. Each Perm object in turn represents a permutation of the architecture graphs vertixes:

>>> p.degree()
4
>>> p
(0, 1)(2, 3)
>>> p[0]
1
>>> p[2]
3
>>> p[4]
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
IndexError: not in domain

We can also directly construct permutation groups and use them in place of architecture graphs via the ArchGraphAutomorphisms class. This can be useful when the automorphism group of an architecture is known ahead of time and there is thus no need to let MPsym determine it:

>>> pg = pympsym.PermGroup(5, ['(0, 1)', '(3, 4)'])
>>> ag = pympsym.ArchGraphAutomorphisms(pg)

A number of convenience methods are available to construct and combine common permutation groups:

>>> pg1 = pympsym.PermGroup.symmetric(5)            # S_5
>>> pg2 = pympsym.PermGroup.cyclic(10)              # C_10
>>> pg = pympsym.PermGroup.direct_product(pg1, pg2) # S_5 x C_10

Limitations

  • Architecture graph initialization could be significantly sped up by employing more advanced BSGS construction algorithms.
  • Heuristic methods could improve accuracy of local search but this would require significant experimentation and fine-tuning work.
  • MPsym contains code for automatically decomposing hierarchical architecture graphs but it is as of now mostly untested and not explicitly useable from either the C++ or Python interface.
  • MPsym also contains code for dealing with partial symmetries derived from [5], however, this is currently broken due to both technical and theoretical problems.

Developer Notes

Main Project Structure

The include and source directories contain the main C++ code which can be compiled into a shared object (or static library if the LINK_STATIC CMake flag is set). The C++ code has four dependencies:

  • Boost
  • Lua
  • nauty
  • nlohmann/json

The former two must already be installed on your system. The latter two are automatically downloaded during the CMake configuration step. nauty is compiled into a separate shared object (or static library), see nauty/CMakeLists.txt.

The lua directory contains the mpsym.lua Lua module which can be used to construct Lua architecture graph description files. When trying to parse these files from MPsym, mpsym.lua must thus be made available via the LUA_PATH environment variable, i.e. by setting:

export LUA_PATH=$LUA_PATH;$(readlink -f mpsym/lua/?.lua)

Alternatively, you can set the LUA_EMBED CMake flag to embed mpsym.lua into the MPsym shared object/static library (which is arguably a weird thing to do but quite handy in practice).

The test/tests directory contains C++ unit tests. these are built by CMake if -DCMAKE_BUILD_TYPE=Debug is specified. The tests use the Googletest framework which is automatically downloaded during the CMake configuration step.

The python directory contains Python binding code and tests. It has the following structure:

python/
├── setup.py
├── mpsym
│   ├── __init__.py
│   └── _mpsym_tests.py
└── source
    ├── CMakeLists.txt
    └── _mpsym.cpp

python/mpsym is the binding module directory. python/_mpsym.cpp contains pybind11 wrapper code for MPsym's public C++ interface. When the PYTHON_BINDINGS CMake flag is set, a corresponding shared object is created under python/mpsym/_mpsym.*.so. _mpsym_tests.py contains a number of unit tests. The mpsym binding module loads both the module created via pybind11 and _mpsym_tests.py in its __init__.py. The latter can be run by invoking mpsym.test(verbosity=...), a return value of 0 indicates success.

If you don't care about the C++ interface, you can directly install the binding module to your system via python setup.py install --user.

Profiling

Besides Release and Debug, a third build mode, Profile, is also supported. In this mode, the programs under profile/source are compiled. They can be used to profile the runtime of the Schreier-Sims algorithm as implemented by MPsym, as well as the various canonical representative algorithms. These programs implement --help flags that should more or less explain how to use them. Some related example architecture graphs and scripts can be found here.

Deploying

Running deploy.sh will create test coverage data and Doxygen documentation and will upload these to Codecov and GitHub pages respectively. Of course you shouldn't be able to do this unless you're me :o).

Continuous Integration

Previously, MPsym used Travis for CI. Since Travis is unfortunately no longer free for FOSS projects, GitHub Actions are now used instead. Take a look at .github/workflows/workflow.yml for details. On every commit and pull request to the master branch, MPsym is built and tested inside a Ubuntu/macOS image using recent versions of gcc/clang.

Previously, Travis also took care of deployment. To save on GitHub Action credits, coverage and documentation must now be deployed manually. PyPi wheels are still built automatically (for Linux and macOS), but now in a separate repository that uses multibuild (currently a work in progress).

References

[1] Holt, D. F. (2005). Handbook of Computational Group Theory. CRC Press

[2] Groups, T. G. (2020). GAP - Groups, Algorithms, and Programming, Version 4.11.0.

[3] McKay, B. D. and Piperno, A. (2014). Practical graph isomorphism, ii. Journal of Symbolic Computation, 60:94–112.

[4] Donaldson, A. F. and Miller, A. (2009). On the constructive orbit problem. Annals of Mathematics and Artificial Intelligence 57:1-35.

[5] East, J., Egri-Nagy, A., Mitchell, J., and Péresse, Y. (2019). Computing finite semigroups. Journal of Symbolic Computation, 92:110–155.