Merge pull request #81 from GraphIt-DSL/yunming

Yunming

README.md

@@ -147,9 +147,16 @@ schedule:
 
 The **test/input** and **test/input\_with\_schedules** directories contain many examples of the algorithm and schedule files. Use them as references when writing your own schedule.
 
+ We provide **more detailed instructions on evaluating the code generation and performance capability** of GraphIt in **graphit/graphit_eval/GraphIt_Evaluation_Guide.md**. In the guide, we provide instructions for using a series of scripts that make it easeir for people to evaluate GraphIt.. 
+
 Input Graph Formats
 ===========
 
 GraphIt reuses [GAPBS input formats](https://github.com/sbeamer/gapbs). Specifically, we have tested with edge list file (.el), weighted edge list file (.wel), binary edge list (.sg), and weighted binary edge list (.wsg) formats. Users can use the converters in GAPBS (GAPBS/src/converter.cc) to convert other graph formats into the supported formats, or convert weighted and unweighted edge list files into their respective binary formats. 
 
 We have provided sample input graph files in the `graphit/test/graphs/` directory. The python tests use the sample input files. 
+
+Autotuning GraphIt Schedules
+===========
+Pleaes refer to **README.md** in **graphit/auotune** for more details. 
+The auotuner is still somehwat experimental. Please read the instructions carefully before trying it out. 

apps/bfs.gt

@@ -23,13 +23,16 @@ func main()
 	startTimer();
         vertices.apply(reset);
     	var frontier : vertexset{Vertex} = new vertexset{Vertex}(0);
-    	frontier.addVertex(14);
-    	parent[14] = 14;
+	var start_vertex : int = atoi(argv[2]);
+    	frontier.addVertex(start_vertex);
+    	parent[start_vertex] = start_vertex;
 
     	while (frontier.getVertexSetSize() != 0)
-            #s1# frontier = edges.from(frontier).to(toFilter).applyModified(updateEdge,parent, true);
+            #s1# var output : vertexset{Vertex} = edges.from(frontier).to(toFilter).applyModified(updateEdge,parent, true);
+	    delete frontier;
+	    frontier = output;
     	end
-
+	delete frontier;
         var elapsed_time : float = stopTimer();
     	print "elapsed time: ";
     	print elapsed_time;

apps/cc.gt

@@ -22,9 +22,11 @@ func main()
         var frontier : vertexset{Vertex} = new vertexset{Vertex}(n);
         vertices.apply(init);
         while (frontier.getVertexSetSize() != 0)
-            #s1# frontier = edges.from(frontier).applyModified(updateEdge,IDs);
+            #s1# var output: vertexset{Vertex} = edges.from(frontier).applyModified(updateEdge,IDs);
+	    delete frontier;
+	    frontier = output;
         end
-
+	delete frontier;
         var elapsed_time : float = stopTimer();
         print "elapsed time: ";
         print elapsed_time;

apps/pagerankdelta.gt

@@ -49,12 +49,16 @@ func main()
 
         for i in 1:11
             #s1# edges.from(frontier).apply(updateEdge);
+            var output : vertexset{Vertex};
             if i == 1
-                frontier = vertices.where(updateVertexFirstRound);
+                output = vertices.where(updateVertexFirstRound);
             else
-                frontier = vertices.where(updateVertex);
+                output = vertices.where(updateVertex);
             end
+        delete frontier;
+        frontier = output;
         end
+	delete frontier;
 
          var elapsed_time : double = stopTimer();
          print "elapsed time: ";

apps/sssp.gt

@@ -19,8 +19,9 @@ func main()
         vertices.apply(reset);
         var n : int = edges.getVertices();
         var frontier : vertexset{Vertex} = new vertexset{Vertex}(0);
-        frontier.addVertex(14); %add source vertex
-        SP[14] = 14;
+	var start_vertex : int = atoi(argv[2]);
+        frontier.addVertex(start_vertex); %add source vertex
+        SP[start_vertex] = 0;
 
         var rounds : int = 0;
         while (frontier.getVertexSetSize() != 0)

autotune/README.md

@@ -0,0 +1,44 @@
+GraphIt Auotuner
+=========
+The autotuner for GraphIt aims to automatically find the best schedules for a given input algorithm specification and input graph. The autotuner is built on top of OpenTuner to efficiently search for the best combination of schedules (optimizations). **Currently, the autotuner still requires the user to place a label "s1" on the operator that nees to be tuned.**
+
+The autotuner is still very experimental. Please read the instructions below carefully before using it. 
+
+Dependencies
+-------------------
+
+
+Please first install [OpenTuner](https://github.com/jansel/opentuner). GraphIt's autotuner is built on top of OpenTuner. 
+
+Autotune Schedules
+-------------------
+
+Currently it is the easiet to use the graphit files under autotune/apps because we need to have hardcoded starting points for the BFS and SSSP. If you want to change the starting point for BFS or SSSP, simply edit bfs_benmark.gt or sssp_benchmark.gt. 
+
+
+To tune the performance of bfs on the small test graph 4.el, use the following command. `--algo_file` specifies the algorithm file, and `--graph` specifies the input graph to tune on. The autotuner have options that enable/disable parallelization, enable/disable NUMA optimizations (by default we disabled NUMA optimization), and other options such as setting an upper bound on the number of segments to use for configNumSSG. We usually put a **time limt** on tuning with the `--stop-after`  option. Even for programs that run on a large input graph, we can usually finish within 5000 seconds. 
+```
+#change into the directory for autotuner
+cd graphit/auotune
+
+#autotune serial pagerank with 4.el graph for 10 seconds
+python graphit_autotuner.py --enable_parallel_tuning 0 --algo_file apps/pagerank_benchmark.gt --graph ../test/graphs/4.el --stop-after 10
+
+#autotune parallel pagerank with 4.el graph for 10 seconds
+python graphit_autotuner.py --enable_parallel_tuning 1 --algo_file apps/pagerank_benchmark.gt --graph ../test/graphs/4.el --stop-after 10
+
+```
+
+To see all the options  
+```
+python graphit_autotuner.py -h
+```
+
+The final result will be displayed as a configuration in the standard output, and also in `final_config.json`. One example is shown below. You can then translate that into the scheduling commands. At this time, this translation step still has to be done manually. 
+```
+('Final Configuration:', {'parallelization': 'serial', 'direction': 'SparsePush', 'numSSG': 5, 'DenseVertexSet': 'boolean-array'})
+```
+
+You can also set the serial and parallel compiler used for the C++ files in the graphit_autotune.py file by modifying the variables `serial_compiler` and `par_compiler`.
+
+Currently, the autotuner can not tune the data layout. So the user would need to provide schedules for fusing together different vectors. The user can provide some scheduling commands in a file as a necessary schedule using the `--default_schedule_file`  option. 

autotune/apps/bfs_benchmark.gt

@@ -24,10 +24,10 @@ func main()
         vertices.apply(reset);
     	var frontier : vertexset{Vertex} = new vertexset{Vertex}(0);
     	frontier.addVertex(14);
-    	parent[14] = 14;
+    	parent[5] = 5;
 
     	while (frontier.getVertexSetSize() != 0)
-            #s1# frontier = edges.from(frontier).to(toFilter).applyModified(updateEdge,parent);
+            #s1# frontier = edges.from(frontier).to(toFilter).applyModified(updateEdge,parent, true);
     	end
 
         var elapsed_time : float = stopTimer();

autotune/graphit_autotuner.py

@@ -13,6 +13,9 @@
 from sys import exit
 import argparse
 
+py_graphitc_file = "../build/bin/graphitc.py"
+serial_compiler = "icc"
+par_compiler = "icpc"
 
 class GraphItTuner(MeasurementInterface):
     new_schedule_file_name = ''
@@ -167,15 +170,15 @@ def compile(self, cfg,  id):
 
 
         #compile the schedule file along with the original algorithm file
-        compile_graphit_cmd = 'python graphitc.py -a  {algo_file} -f {schedule_file} -i ../include/ -l ../build/lib/libgraphitlib.a  -o test.cpp'.format(algo_file=self.args.algo_file, schedule_file=self.new_schedule_file_name) 
+        compile_graphit_cmd = 'python ' + py_graphitc_file +  ' -a  {algo_file} -f {schedule_file} -i ../include/ -l ../build/lib/libgraphitlib.a  -o test.cpp'.format(algo_file=self.args.algo_file, schedule_file=self.new_schedule_file_name) 
 
         if not self.use_NUMA:
             if not self.enable_parallel_tuning:
                 # if parallel icpc compiler is not needed (only tuning serial schedules)
-                compile_cpp_cmd = 'g++ -std=c++11  -I ../src/runtime_lib/ -O3  test.cpp -o test'
+                compile_cpp_cmd = serial_compiler + ' -std=c++11  -I ../src/runtime_lib/ -O3  test.cpp -o test'
             else:
                 # if parallel icpc compiler is supported and needed
-                compile_cpp_cmd = 'icpc -std=c++11 -DCILK  -I ../src/runtime_lib/ -O3  test.cpp -o test'
+                compile_cpp_cmd = par_compiler + ' -std=c++11 -DCILK  -I ../src/runtime_lib/ -O3  test.cpp -o test'
         else:
             #add the additional flags for NUMA
             compile_cpp_cmd = 'icpc -std=c++11 -DOPENMP -lnuma -DNUMA -qopenmp -I ../src/runtime_lib/ -O3  test.cpp -o test'
@@ -307,7 +310,7 @@ def save_final_config(self, configuration):
     parser = argparse.ArgumentParser(parents=opentuner.argparsers())
     parser.add_argument('--graph', type=str, default="../test/graphs/4.sg",
                     help='the graph to tune on')
-    parser.add_argument('--enable_NUMA_tuning', type=int, default=1, help='enable tuning NUMA-aware schedules. 1 for enable (default), 0 for disable')
+    parser.add_argument('--enable_NUMA_tuning', type=int, default=0, help='enable tuning NUMA-aware schedules. 1 for enable (default), 0 for disable')
     parser.add_argument('--enable_parallel_tuning', type=int, default=1, help='enable tuning paralleliation schedules. 1 for enable (default), 0 for disable')
     parser.add_argument('--enable_denseVertexSet_tuning', type=int, default=1, help='enable tuning denseVertexSet schedules. 1 for enable (default), 0 for disable')
     parser.add_argument('--algo_file', type=str, required=True, help='input algorithm file')