gwdidonato / hpgda-spring22 Goto Github PK

View Code? Open in Web Editor NEW

8.0 3.0 10.0 64.43 MB

Jupyter Notebook 53.08% Makefile 1.02% Shell 0.46% Cuda 21.11% C 6.01% C++ 18.20% Euphoria 0.09% Verilog 0.01%

hpgda-spring22's Introduction

High-Performance Graph & Data Analytics - Spring 2022

Repository for the course High-Performance Graph & Data Analytics at Polimi, Spring 2022

cuda_CONTEST contains the source code for the GPU track of the contest. See the readme in the folder for more information.
data_struct_CONTEST contains the source code for the graph data structures track of the contest. See the readme in the folder for more information.
graph_ml_DEMO contains a python notebook and the input data for a demo where graph machine learning is used to predict links in a protein-protein interaction network. See the readme in the folder for more information.

hpgda-spring22's People

Contributors

Stargazers

Watchers

Forkers

calonca fedann andreaortenzi mattiataglia andreabosisio alessiobraccini ottaviabelotti pierorendina paolo6galli

hpgda-spring22's Issues

PPR Benchmark not compliant with the contest rules

In the PPR benchmark file, the precision of the results is evaluated through 2 unordered sets.

hpgda-spring22/cuda_CONTEST/src/benchmarks/personalized_pagerank.cu

Lines 143 to 144 in 0b1b250

 std::unordered_set<int> top_pr_indices; 

 std::unordered_set<int> top_pr_golden_indices;

In particular here is where the evaluation happens.

hpgda-spring22/cuda_CONTEST/src/benchmarks/personalized_pagerank.cu

Lines 163 to 168 in 0b1b250

 std::cout.precision(old_precision); 

 // Set intersection to find correctly retrieved vertices; 

 std::vector<int> correctly_retrieved_vertices; 

 set_intersection(top_pr_indices.begin(), top_pr_indices.end(), top_pr_golden_indices.begin(), top_pr_golden_indices.end(), std::back_inserter(correctly_retrieved_vertices)); 

 precision = double(correctly_retrieved_vertices.size()) / topk; 

 if (debug) std::cout << "correctly retrived top-" << topk << " vertices=" << correctly_retrieved_vertices.size() << " (" << 100 * precision << "%)" << std::endl;

The problem I encountered is related to the set_intersection method, which, in this case, to work properly should receive 2 ordered sets (standard c++ sets), because set_intersection expects both sets sorted, otherwise the intersection takes into account the "raw" order of the topK nodes, differently from what stated in the rules.

Here are the results obtained using unordered_set:

Using sets no wrong precision computation occurs.

Possible inconsistency in spmv_coo_cpu

Hi, is it possible that here at line 39 x and y should be swapped when the read_tuples_transposed flag is set to true? According me and @feDann they should be swapped because when the flag is set to true, x holds the column ids and y holds the row ids.

hpgda-spring22/cuda_CONTEST/src/benchmarks/personalized_pagerank.cuh

Lines 37 to 41 in 2416e55

 inline void spmv_coo_cpu(const int *x, const int *y, const double *val, const double *vec, double *result, int N) { 

 for (int i = 0; i < N; i++) { 

 result[x[i]] += val[i] * vec[y[i]]; 

 } 

 }

hpgda-spring22/cuda_CONTEST/src/benchmarks/personalized_pagerank.cu

Lines 51 to 58 in 2416e55

 read_mtx(graph_file_path.c_str(), &x, &y, &val, 

 &num_rows, &num_columns, &E, // Store the number of vertices (row and columns must be the same value), and edges; 

 true, // If true, read edges TRANSPOSED, i.e. edge (2, 3) is loaded as (3, 2). We set this true as it simplifies the PPR computation; 

 false, // If true, read the third column of the matrix file. If false, set all values to 1 (this is what you want when reading a graph topology); 

 debug, 

 false, // MTX files use indices starting from 1. If for whatever reason your MTX files uses indices that start from 0, set zero_indexed_file=true; 

 true // If true, sort the edges in (x, y) order. If you have a sorted MTX file, turn this to false to make loading faster; 

 );

[CUDA][BUG] Precision not correctly updated when debug=false

HI, i noticed that when i run the executable without the -d flag even if the -c flag is present the cpu_validation function is never called so the gpu_result column of the csv is always to zero. This happens only for the ppr implementation because the actual value of the precision is updated inside the cpu_validation function

hpgda-spring22/cuda_CONTEST/src/benchmarks/personalized_pagerank.cu

Line 167 in 2416e55

precision = double(correctly_retrieved_vertices.size()) / topk;

A possible solution could be to move line 84 outside the if statement and also modify the cpu_validation function to print line 136 only if debug=true:

hpgda-spring22/cuda_CONTEST/src/benchmark.cu

Lines 81 to 88 in 2416e55

 if (debug) { 

 std::cout << " GPU execution(" << i << ")=" << float(exec_time) / 1000 << " ms" << std::endl; 

 std::cout << " GPU result=" << print_result() << std::endl; 

 if (do_cpu_validation) cpu_validation(i); 

 } else { 

 std::cout << i << "," << print_result(true) << "," << float(reset_time + exec_time) / 1e6 << "," << float(reset_time) / 1e6 << "," << float(exec_time) / 1e6 << std::endl; 

 } 

 }

hpgda-spring22/cuda_CONTEST/src/benchmarks/personalized_pagerank.cu

Line 136 in 2416e55

 std::cout << "exec time CPU=" << double(exec_time) / 1000 << " ms" << std::endl; 

gwdidonato / hpgda-spring22 Goto Github PK

hpgda-spring22's Introduction

High-Performance Graph & Data Analytics - Spring 2022

hpgda-spring22's People

Contributors

Stargazers

Watchers

Forkers

hpgda-spring22's Issues

PPR Benchmark not compliant with the contest rules

Possible inconsistency in spmv_coo_cpu

[CUDA][BUG] Precision not correctly updated when debug=false

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	std::unordered_set<int> top_pr_indices;
	std::unordered_set<int> top_pr_golden_indices;

	std::cout.precision(old_precision);
	// Set intersection to find correctly retrieved vertices;
	std::vector<int> correctly_retrieved_vertices;
	set_intersection(top_pr_indices.begin(), top_pr_indices.end(), top_pr_golden_indices.begin(), top_pr_golden_indices.end(), std::back_inserter(correctly_retrieved_vertices));
	precision = double(correctly_retrieved_vertices.size()) / topk;
	if (debug) std::cout << "correctly retrived top-" << topk << " vertices=" << correctly_retrieved_vertices.size() << " (" << 100 * precision << "%)" << std::endl;

	inline void spmv_coo_cpu(const int x, const int y, const double val, const double vec, double *result, int N) {
	for (int i = 0; i < N; i++) {
	result[x[i]] += val[i] * vec[y[i]];
	}
	}

	read_mtx(graph_file_path.c_str(), &x, &y, &val,
	&num_rows, &num_columns, &E, // Store the number of vertices (row and columns must be the same value), and edges;
	true, // If true, read edges TRANSPOSED, i.e. edge (2, 3) is loaded as (3, 2). We set this true as it simplifies the PPR computation;
	false, // If true, read the third column of the matrix file. If false, set all values to 1 (this is what you want when reading a graph topology);
	debug,
	false, // MTX files use indices starting from 1. If for whatever reason your MTX files uses indices that start from 0, set zero_indexed_file=true;
	true // If true, sort the edges in (x, y) order. If you have a sorted MTX file, turn this to false to make loading faster;
	);

	if (debug) {
	std::cout << " GPU execution(" << i << ")=" << float(exec_time) / 1000 << " ms" << std::endl;
	std::cout << " GPU result=" << print_result() << std::endl;
	if (do_cpu_validation) cpu_validation(i);
	} else {
	std::cout << i << "," << print_result(true) << "," << float(reset_time + exec_time) / 1e6 << "," << float(reset_time) / 1e6 << "," << float(exec_time) / 1e6 << std::endl;
	}
	}