GraphLab: Distributed Graph-Parallel API
2.1
|
The asynchronous consistent engine executed vertex programs asynchronously but ensures mutual exclusion such that adjacent vertices are never executed simultaneously. Mutual exclusion can be weakened to "factorized" consistency in which case only individual gathers/applys/ scatters are guaranteed to be consistent. More...
#include <graphlab/engine/async_consistent_engine.hpp>
Public Types | |
typedef VertexProgram | vertex_program_type |
The user defined vertex program type. Equivalent to the VertexProgram template argument. | |
typedef VertexProgram::gather_type | gather_type |
The user defined type returned by the gather function. | |
typedef VertexProgram::message_type | message_type |
The user defined message type used to signal neighboring vertex programs. | |
typedef VertexProgram::vertex_data_type | vertex_data_type |
The type of data associated with each vertex in the graph. | |
typedef VertexProgram::edge_data_type | edge_data_type |
The type of data associated with each edge in the graph. | |
typedef VertexProgram::graph_type | graph_type |
The type of graph supported by this vertex program. | |
typedef graph_type::vertex_type | vertex_type |
The type used to represent a vertex in the graph. See graphlab::distributed_graph::vertex_type for details. | |
typedef graph_type::edge_type | edge_type |
The type used to represent an edge in the graph. See graphlab::distributed_graph::edge_type for details. | |
typedef icontext< graph_type, gather_type, message_type > | icontext_type |
The type of the callback interface passed by the engine to vertex programs. See graphlab::icontext for details. | |
typedef graph_type::vertex_id_type | vertex_id_type |
The vertex identifier type defined in graphlab::vertex_id_type. |
Public Member Functions | |
async_consistent_engine (distributed_control &dc, graph_type &graph, const graphlab_options &opts=graphlab_options()) | |
size_t | num_updates () const |
Compute the total number of updates (calls to apply) executed since start was last invoked. | |
float | elapsed_seconds () const |
Get the elapsed time in seconds since start was last called. | |
int | iteration () const |
Not meaningful for the asynchronous engine. Returns -1. | |
void | signal (vertex_id_type gvid, const message_type &message=message_type()) |
void | signal_all (const message_type &message=message_type(), const std::string &order="shuffle") |
void | signal_vset (const vertex_set &vset, const message_type &message=message_type(), const std::string &order="shuffle") |
execution_status::status_enum | start () |
Start the engine execution. | |
aggregator_type * | get_aggregator () |
| |
virtual void | signal (vertex_id_type vertex, const message_type &message=message_type())=0 |
Signals single a vertex with an optional message. | |
virtual void | signal_all (const message_type &message=message_type(), const std::string &order="shuffle")=0 |
Signal all vertices with a particular message. | |
virtual void | signal_vset (const vertex_set &vset, const message_type &message=message_type(), const std::string &order="shuffle")=0 |
Signal a set of vertices with a particular message. | |
template<typename ReductionType , typename VertexMapType , typename FinalizerType > | |
bool | add_vertex_aggregator (const std::string &key, VertexMapType map_function, FinalizerType finalize_function) |
Creates a vertex aggregator. Returns true on success. Returns false if an aggregator of the same name already exists. | |
template<typename ReductionType , typename EdgeMapType , typename FinalizerType > | |
bool | add_edge_aggregator (const std::string &key, EdgeMapType map_function, FinalizerType finalize_function) |
Creates an edge aggregator. Returns true on success. Returns false if an aggregator of the same name already exists. | |
bool | aggregate_now (const std::string &key) |
Performs an immediate aggregation on a key. | |
template<typename ReductionType , typename VertexMapperType > | |
ReductionType | map_reduce_vertices (VertexMapperType mapfunction) |
Performs a map-reduce operation on each vertex in the graph returning the result. | |
template<typename ReductionType , typename EdgeMapperType > | |
ReductionType | map_reduce_edges (EdgeMapperType mapfunction) |
Performs a map-reduce operation on each edge in the graph returning the result. | |
template<typename VertexMapperType > | |
void | transform_vertices (VertexMapperType mapfunction) |
Performs a transformation operation on each vertex in the graph. | |
template<typename EdgeMapperType > | |
void | transform_edges (EdgeMapperType mapfunction) |
Performs a transformation operation on each edge in the graph. | |
bool | aggregate_periodic (const std::string &key, float seconds) |
Requests that a particular aggregation key be recomputed periodically when the engine is running. |
The asynchronous consistent engine executed vertex programs asynchronously but ensures mutual exclusion such that adjacent vertices are never executed simultaneously. Mutual exclusion can be weakened to "factorized" consistency in which case only individual gathers/applys/ scatters are guaranteed to be consistent.
VertexProgram | The user defined vertex program type which should implement the graphlab::ivertex_program interface. |
On start() the graphlab::ivertex_program::init function is invoked on all vertex programs in parallel to initialize the vertex program, vertex data, and possibly signal vertices.
After which, the engine spawns a collection of threads where each thread individually performs the following tasks:
The engine threads multiplexes the above procedure through a secondary internal queue, allowing an arbitrary large number of vertices to begin processing at the same time.
The asynchronous consistent engine is constructed by passing in a graphlab::distributed_control object which manages coordination between engine threads and a graphlab::distributed_graph object which is the graph on which the engine should be run. The graph should already be populated and cannot change after the engine is constructed. In the distributed setting all program instances (running on each machine) should construct an instance of the engine at the same time.
Computation is initiated by signaling vertices using either graphlab::async_consistent_engine::signal or graphlab::async_consistent_engine::signal_all. In either case all machines should invoke signal or signal all at the same time. Finally, computation is initiated by calling the graphlab::async_consistent_engine::start function.
The following is a simple example demonstrating how to use the engine:
The asynchronous engine supports several engine options which can be set as command line arguments using –engine_opts
:
Definition at line 208 of file async_consistent_engine.hpp.
typedef VertexProgram::edge_data_type graphlab::async_consistent_engine< VertexProgram >::edge_data_type |
The type of data associated with each edge in the graph.
The edge data type must be Serializable.
Definition at line 255 of file async_consistent_engine.hpp.
typedef graph_type::edge_type graphlab::async_consistent_engine< VertexProgram >::edge_type |
The type used to represent an edge in the graph. See graphlab::distributed_graph::edge_type for details.
The edge type contains the function graphlab::distributed_graph::edge_type::data which returns a reference to the edge data. In addition the edge type contains the function graphlab::distributed_graph::edge_type::source and graphlab::distributed_graph::edge_type::target.
Reimplemented from graphlab::iengine< VertexProgram >.
Definition at line 288 of file async_consistent_engine.hpp.
typedef VertexProgram::gather_type graphlab::async_consistent_engine< VertexProgram >::gather_type |
The user defined type returned by the gather function.
The gather type is defined in the graphlab::ivertex_program interface and is the value returned by the graphlab::ivertex_program::gather function. The gather type must have an operator+=(const gather_type& other)
function and must be Serializable.
Definition at line 229 of file async_consistent_engine.hpp.
typedef VertexProgram::graph_type graphlab::async_consistent_engine< VertexProgram >::graph_type |
The type of graph supported by this vertex program.
See graphlab::distributed_graph
Reimplemented from graphlab::iengine< VertexProgram >.
Definition at line 262 of file async_consistent_engine.hpp.
typedef icontext<graph_type, gather_type, message_type> graphlab::async_consistent_engine< VertexProgram >::icontext_type |
The type of the callback interface passed by the engine to vertex programs. See graphlab::icontext for details.
The context callback is passed to the vertex program functions and is used to signal other vertices, get the current iteration, and access information about the engine.
Reimplemented from graphlab::iengine< VertexProgram >.
Definition at line 298 of file async_consistent_engine.hpp.
typedef VertexProgram::message_type graphlab::async_consistent_engine< VertexProgram >::message_type |
The user defined message type used to signal neighboring vertex programs.
The message type is defined in the graphlab::ivertex_program interface and used in the call to graphlab::icontext::signal. The message type must have an operator+=(const gather_type& other)
function and must be Serializable.
Reimplemented from graphlab::iengine< VertexProgram >.
Definition at line 241 of file async_consistent_engine.hpp.
typedef VertexProgram::vertex_data_type graphlab::async_consistent_engine< VertexProgram >::vertex_data_type |
The type of data associated with each vertex in the graph.
The vertex data type must be Serializable.
Definition at line 248 of file async_consistent_engine.hpp.
typedef VertexProgram graphlab::async_consistent_engine< VertexProgram >::vertex_program_type |
The user defined vertex program type. Equivalent to the VertexProgram template argument.
The user defined vertex program type which should implement the graphlab::ivertex_program interface.
Reimplemented from graphlab::iengine< VertexProgram >.
Definition at line 218 of file async_consistent_engine.hpp.
typedef graph_type::vertex_type graphlab::async_consistent_engine< VertexProgram >::vertex_type |
The type used to represent a vertex in the graph. See graphlab::distributed_graph::vertex_type for details.
The vertex type contains the function graphlab::distributed_graph::vertex_type::data which returns a reference to the vertex data as well as other functions like graphlab::distributed_graph::vertex_type::num_in_edges which returns the number of in edges.
Reimplemented from graphlab::iengine< VertexProgram >.
Definition at line 275 of file async_consistent_engine.hpp.
|
inline |
Constructs an asynchronous consistent distributed engine. The number of threads to create are read from opts.get_ncpus(). The scheduler to construct is read from opts.get_scheduler_type(). The default scheduler is the queued_fifo scheduler. For details on the scheduler types
See the main class documentation for the available engine options.
dc | Distributed controller to associate with |
graph | The graph to schedule over. The graph must be fully constructed and finalized. |
opts | A graphlab::graphlab_options object containing options and parameters for the scheduler and the engine. |
Definition at line 669 of file async_consistent_engine.hpp.
|
inlineinherited |
Creates an edge aggregator. Returns true on success. Returns false if an aggregator of the same name already exists.
Creates a edge aggregator associated to a particular key. The map_function is called over every edge in the graph, and the return value of the map is summed. The finalize_function is then called on the result of the reduction. The finalize_function is called on all machines. The map_function should only read the graph data, and should not make any modifications.
For instance, if the graph has float vertex data, and float edge data:
An aggregator can be constructed to compute the absolute sum of all the edge data. To do this, we define two functions.
Next, we define the aggregator in the engine by calling add_edge_aggregator(). We must assign it a unique name which will be used to reference this particular aggregate operation. We shall call it "absolute_edge_sum".
When executed, the engine execute absolute_edge_data()
on each edge in the graph. absolute_edge_data()
reads the edge data, and returns its absolute value. All return values are then summing them together using the float's += operator. The final result is than passed to the print_finalize
function. The template argument <float>
is necessary to provide information about the return type of absolute_edge_data
.
This aggregator can be run immediately by calling aggregate_now() with the name of the aggregator.
Or can be arranged to run periodically together with the engine execution (in this example, every 1.5 seconds).
Note that since finalize is called on all machines, multiple copies of the total will be printed. If only one copy is desired, see context.cout() or to get the actual process ID using context.procid()
The add_edge_aggregator() function is also templatized over both function types and there is no strong enforcement of the exact argument types of the map function and the reduce function. For instance, in the above example, the following print_finalize() variants may also be accepted.
In particlar, the last variation may be useful for performance reasons if the reduction type is large.
To obtain consistent distributed behavior in the distributed setting, we designed the aggregator to minimize the amount of asymmetry among the machines. In particular, the finalize operation is guaranteed to be called on all machines. This therefore permits global variables to be modified on finalize since all machines are ensured to be eventually consistent.
For instance, in the above example, print_finalize could store the result in a global variable:
which will make it accessible to all other running update functions.
ReductionType | The output of the map function. Must have operator+= defined, and must be Serializable. |
EdgeMapperType | The type of the map function. Not generally needed. Can be inferred by the compiler. |
FinalizerType | The type of the finalize function. Not generally needed. Can be inferred by the compiler. |
[in] | key | The name of this aggregator. Must be unique. |
[in] | map_function | The Map function to use. Must take an icontext_type& as its first argument, and a edge_type, or a reference to a edge_type as its second argument. Returns a ReductionType which must be summable and Serializable . |
[in] | finalize_function | The Finalize function to use. Must take an icontext_type& as its first argument and a ReductionType, or a reference to a ReductionType as its second argument. |
Definition at line 692 of file iengine.hpp.
|
inlineinherited |
Creates a vertex aggregator. Returns true on success. Returns false if an aggregator of the same name already exists.
Creates a vertex aggregator associated to a particular key. The map_function is called over every vertex in the graph, and the return value of the map is summed. The finalize_function is then called on the result of the reduction. The finalize_function is called on all machines. The map_function should only read the graph data, and should not make any modifications.
For instance, if the graph has float vertex data, and float edge data:
An aggregator can be constructed to compute the absolute sum of all the vertex data. To do this, we define two functions.
Next, we define the aggregator in the engine by calling add_vertex_aggregator(). We must assign it a unique name which will be used to reference this particular aggregate operation. We shall call it "absolute_vertex_sum".
When executed, the engine execute absolute_vertex_data()
on each vertex in the graph. absolute_vertex_data()
reads the vertex data, and returns its absolute value. All return values are then summing them together using the float's += operator. The final result is than passed to the print_finalize
function. The template argument <float>
is necessary to provide information about the return type of absolute_vertex_data
.
This aggregator can be run immediately by calling aggregate_now() with the name of the aggregator.
Or can be arranged to run periodically together with the engine execution (in this example, every 1.5 seconds).
Note that since finalize is called on all machines, multiple copies of the total will be printed. If only one copy is desired, see context.cout() or to get the actual process ID using context.procid()
In practice, the reduction type can be any arbitrary user-defined type as long as a += operator is defined. This permits great flexibility in the type of operations the aggregator can perform.
The add_vertex_aggregator() function is also templatized over both function types and there is no strong enforcement of the exact argument types of the map function and the reduce function. For instance, in the above example, the following print_finalize() variants may also be accepted.
In particlar, the last variation may be useful for performance reasons if the reduction type is large.
To obtain consistent distributed behavior in the distributed setting, we designed the aggregator to minimize the amount of asymmetry among the machines. In particular, the finalize operation is guaranteed to be called on all machines. This therefore permits global variables to be modified on finalize since all machines are ensured to be eventually consistent.
For instance, in the above example, print_finalize could store the result in a global variable:
which will make it accessible to all other running update functions.
ReductionType | The output of the map function. Must have operator+= defined, and must be Serializable. |
VertexMapperType | The type of the map function. Not generally needed. Can be inferred by the compiler. |
FinalizerType | The type of the finalize function. Not generally needed. Can be inferred by the compiler. |
[in] | map_function | The Map function to use. Must take an |
[in] | key | The name of this aggregator. Must be unique. icontext_type& as its first argument, and a vertex_type, or a reference to a vertex_type as its second argument. Returns a ReductionType which must be summable and Serializable . |
[in] | finalize_function | The Finalize function to use. Must take an icontext_type& as its first argument and a ReductionType, or a reference to a ReductionType as its second argument. |
Definition at line 484 of file iengine.hpp.
|
inlineinherited |
Performs an immediate aggregation on a key.
Performs an immediate aggregation on a key. All machines must call this simultaneously. If the key is not found, false is returned. Otherwise returns true on success.
For instance, the following code will run the aggregator with the name "absolute_vertex_sum" immediately.
[in] | key | Key to aggregate now. Must be a key previously created by add_vertex_aggregator() or add_edge_aggregator(). |
Definition at line 780 of file iengine.hpp.
|
inlineinherited |
Requests that a particular aggregation key be recomputed periodically when the engine is running.
Requests that the aggregator with a given key be aggregated every certain number of seconds when the engine is running. Note that the period is prescriptive: in practice the actual period will be larger than the requested period. Seconds must be >= 0;
For instance, the following code will schedule the aggregator with the name "absolute_vertex_sum" to run every 1.5 seconds.
[in] | key | Key to schedule. Must be a key previously created by add_vertex_aggregator() or add_edge_aggregator(). |
[in] | seconds | How frequently to schedule. Must be >= 0. seconds == 0 will ensure that this key is continously recomputed. |
All machines must call simultaneously.
Definition at line 1169 of file iengine.hpp.
|
inlinevirtual |
Get the elapsed time in seconds since start was last called.
Implements graphlab::iengine< VertexProgram >.
Definition at line 851 of file async_consistent_engine.hpp.
|
inline |
Definition at line 2352 of file async_consistent_engine.hpp.
|
inlineinherited |
Performs a map-reduce operation on each edge in the graph returning the result.
Given a map function, map_reduce_edges() call the map function on all edges in the graph. The return values are then summed together and the final result returned. The map function should only read data and should not make any modifications. map_reduce_edges() must be called on all machines simultaneously.
For instance, if the graph has float vertex data, and float edge data:
To compute an absolute sum over all the edge data, we would write a function which reads in each a edge, and returns the absolute value of the data on the edge.
After which calling:
will call the absolute_edge_data()
function on each edge in the graph. absolute_edge_data()
reads the value of the edge and returns the absolute result. This return values are then summed together and returned. All machines see the same result.
The template argument <float>
is needed to inform the compiler regarding the return type of the mapfunction.
Another common use for the map_reduce_edges() function is in signalling. Since the map function is passed a context, it can be used to perform signalling of edges for execution during a later engine.start() call.
For instance, the following code will signal the source vertex of each edge.
Note that in this case, we are not interested in a reduction operation, and thus we return a graphlab::empty object. Calling:
will run signal_source()
on all edges, signalling all source vertices.
The map function has the same structure as that in add_edge_aggregator() and may be reused in an aggregator. This function is also very similar to graphlab::distributed_graph::map_reduce_edges() with the difference that this takes a context and thus can be used to perform signalling. Finally transform_edges() can be used to perform a similar but may also make modifications to graph data.
ReductionType | The output of the map function. Must have operator+= defined, and must be Serializable. |
EdgeMapperType | The type of the map function. Not generally needed. Can be inferred by the compiler. |
mapfunction | The map function to use. Must take an icontext_type& as its first argument, and a edge_type, or a reference to a edge_type as its second argument. Returns a ReductionType which must be summable and Serializable . |
Definition at line 974 of file iengine.hpp.
|
inlineinherited |
Performs a map-reduce operation on each vertex in the graph returning the result.
Given a map function, map_reduce_vertices() call the map function on all vertices in the graph. The return values are then summed together and the final result returned. The map function should only read the vertex data and should not make any modifications. map_reduce_vertices() must be called on all machines simultaneously.
For instance, if the graph has float vertex data, and float edge data:
To compute an absolute sum over all the vertex data, we would write a function which reads in each a vertex, and returns the absolute value of the data on the vertex.
After which calling:
will call the absolute_vertex_data()
function on each vertex in the graph. absolute_vertex_data()
reads the value of the vertex and returns the absolute result. This return values are then summed together and returned. All machines see the same result.
The template argument <float>
is needed to inform the compiler regarding the return type of the mapfunction.
Another common use for the map_reduce_vertices() function is in signalling. Since the map function is passed a context, it can be used to perform signalling of vertices for execution during a later engine.start() call.
For instance, the following code will signal all vertices with value >= 1
Note that in this case, we are not interested in a reduction operation, and thus we return a graphlab::empty object. Calling:
will run signal_vertices()
on all vertices, signalling all vertices with value <= 1
The map function has the same structure as that in add_vertex_aggregator() and may be reused in an aggregator. This function is also very similar to graphlab::distributed_graph::map_reduce_vertices() with the difference that this takes a context and thus can be used to perform signalling. Finally transform_vertices() can be used to perform a similar but may also make modifications to graph data.
ReductionType | The output of the map function. Must have operator+= defined, and must be Serializable. |
VertexMapperType | The type of the map function. Not generally needed. Can be inferred by the compiler. |
mapfunction | The map function to use. Must take an icontext_type& as its first argument, and a vertex_type, or a reference to a vertex_type as its second argument. Returns a ReductionType which must be summable and Serializable . |
Definition at line 876 of file iengine.hpp.
|
inlinevirtual |
Compute the total number of updates (calls to apply) executed since start was last invoked.
Implements graphlab::iengine< VertexProgram >.
Definition at line 842 of file async_consistent_engine.hpp.
|
pure virtualinherited |
Signals single a vertex with an optional message.
This function sends a message to particular vertex which will receive that message on start. The signal function must be invoked on all machines simultaneously. For example:
and not:
Since signal is executed synchronously on all machines it should only be used to schedule a small set of vertices. The preferred method to signal a large set of vertices (e.g., all vertices that are a certain type) is to use either the vertex program init function or the aggregation framework. For example to signal all vertices that have a particular value one could write:
[in] | vid | the vertex id to signal |
[in] | message | the message to send to that vertex. The default message is sent if no message is provided. (See ivertex_program::message_type for details about the message_type). |
Implemented in graphlab::semi_synchronous_engine< VertexProgram >, and graphlab::omni_engine< VertexProgram >.
|
pure virtualinherited |
Signal all vertices with a particular message.
This function sends the same message to all vertices which will receive that message on start. The signal_all function must be invoked on all machines simultaneously. For example:
and not:
The signal_all function is the most common way to send messages to the engine. For example in the pagerank application we want all vertices to be active on the first round. Therefore we would write:
[in] | message | the message to send to all vertices. The default message is sent if no message is provided (See ivertex_program::message_type for details about the message_type). |
Implemented in graphlab::semi_synchronous_engine< VertexProgram >, and graphlab::omni_engine< VertexProgram >.
|
pure virtualinherited |
Signal a set of vertices with a particular message.
This function sends the same message to a set of vertices which will receive that message on start. The signal_vset function must be invoked on all machines simultaneously. For example:
signal_all() is conceptually equivalent to:
[in] | vset | The set of vertices to signal |
[in] | message | the message to send to all vertices. The default message is sent if no message is provided (See ivertex_program::message_type for details about the message_type). |
Implemented in graphlab::semi_synchronous_engine< VertexProgram >, and graphlab::omni_engine< VertexProgram >.
|
inlinevirtual |
Start the engine execution.
This function starts the engine and does not return until the scheduler has no tasks remaining.
Implements graphlab::iengine< VertexProgram >.
Definition at line 2163 of file async_consistent_engine.hpp.
|
inlineinherited |
Performs a transformation operation on each edge in the graph.
Given a mapfunction, transform_edges() calls mapfunction on every edge in graph. The map function may make modifications to the data on the edge. transform_edges() must be called on all machines simultaneously.
For instance, if the graph has integer vertex data, and integer edge data:
To set each edge value to be the number of out-going edges of the target vertex, we may write the following:
Calling transform_edges():
will run the set_edge_value()
function on each edge in the graph, setting its new value.
Since the mapfunction is provided with a context, the mapfunction can also be used to perform signalling. For instance, the set_edge_value
function above may be modified to set the value of the edge, but to also signal the target vertex.
However, if the purpose of the function is to only signal without making modifications, map_reduce_edges() will be more efficient as this function will additionally perform distributed synchronization of modified data.
map_reduce_edges() provide similar signalling functionality, but should not make modifications to graph data. graphlab::distributed_graph::transform_edges() provide the same graph modification capabilities, but without a context and thus cannot perform signalling.
EdgeMapperType | The type of the map function. Not generally needed. Can be inferred by the compiler. |
mapfunction | The map function to use. Must take an icontext_type& as its first argument, and a edge_type, or a reference to a edge_type as its second argument. Returns void. |
Definition at line 1132 of file iengine.hpp.
|
inlineinherited |
Performs a transformation operation on each vertex in the graph.
Given a mapfunction, transform_vertices() calls mapfunction on every vertex in graph. The map function may make modifications to the data on the vertex. transform_vertices() must be called by all machines simultaneously.
For instance, if the graph has integer vertex data, and integer edge data:
To set each vertex value to be the number of out-going edges, we may write the following function:
Calling transform_vertices():
will run the set_vertex_value()
function on each vertex in the graph, setting its new value.
Since the mapfunction is provided with a context, the mapfunction can also be used to perform signalling. For instance, the set_vertex_value
function above may be modified to set the value of the vertex, but to also signal the vertex if it has more than 5 outgoing edges.
However, if the purpose of the function is to only signal without making modifications, map_reduce_vertices() will be more efficient as this function will additionally perform distributed synchronization of modified data.
map_reduce_vertices() provide similar signalling functionality, but should not make modifications to graph data. graphlab::distributed_graph::transform_vertices() provide the same graph modification capabilities, but without a context and thus cannot perform signalling.
VertexMapperType | The type of the map function. Not generally needed. Can be inferred by the compiler. |
mapfunction | The map function to use. Must take an icontext_type& as its first argument, and a vertex_type, or a reference to a vertex_type as its second argument. Returns void. |
Definition at line 1055 of file iengine.hpp.