chenxuhao / graphminer Goto Github PK

Hello,

I'm having an issue with the GPU code, when I run even with the graph provided I always get zero for the number of triangles or other patterns. The OpenMP code seems to work fine with me. I provided the output of both GPU and OpenMP.

GPU:

./tc_gpu_base /home/ghadeer/cit-Patents/graph
Triangle Counting (for undirected graphs only)
Using DAG (static orientation)
Orientation enabled, using DAG
Time on generating the DAG: 0.216688 sec
|V| 3774769 |E| 16518947
Found 4 devices
Device[0]: NVIDIA TITAN V
Compute capability: 7.0
Warp size: 32
Total # SM: 80
Total # CUDA cores: 5120
Total amount of shared memory per block: 49152 bytes
Total # registers per block: 65536
Total amount of constant memory: 65536 bytes
Total global memory: 11.8 GB
Memory Clock Rate: 0.81 GHz
Memory Bus Width: 3072 bits
Peak Memory Bandwidth: 652.80 GB/s

Device[1]: NVIDIA GeForce RTX 2070 SUPER
Device[2]: NVIDIA TITAN Xp
Device[3]: Tesla K40c
GPU_total_mem = 12652838912 graph_mem = 162349736
Time on generating the edgelist: 0.070154 sec
Edge parallel: edgelist size = 16518947
max_blocks_per_SM = 0
CUDA triangle counting (65536 CTAs, 256 threads/CTA)
runtime [gpu_base] = 2e-06 sec
throughput = 8259.47 billion Traversed Edges Per Second (TEPS)
total_num_triangles = 0

OpenMP (which seems to work fine)

./tc_gpu_base 0 /home/ghadeer/cit-Patents/graph
Triangle Counting (for undirected graphs only)
Using DAG (static orientation)
tc_gpu_base: ../common/graph.cc:8: Graph::Graph(std::string, bool, bool): Assertion `f_meta' failed.
Aborted (core dumped)
[gaa54@ithaca bin]$ ./tc_gpu_base
Usage: ./tc_gpu_base [num_gpu(1)] [chunk_size(1024)]
Example: ./tc_gpu_base /graph_inputs/mico/graph
[gaa54@ithaca bin]$ ./tc_omp_base /home/ghadeer/cit-Patents/graph 0
Triangle Counting (for undirected graphs only)
Using DAG (static orientation)
Orientation enabled, using DAG
Time on generating the DAG: 0.215716 sec
|V| 3774769 |E| 16518947
Launching OpenMP TC solver (32 threads) ...
runtime [omp_base] = 0.241254 sec
total_num_triangles = 7515023

Thank you

Compiling fsm with cuda 12 gives following error with the code.

g++ -Wall -fopenmp -std=c++17 -march=native -O3 -I../../include -I./ -I./dfscode -c omp_base.cc
g++ -Wall -fopenmp -std=c++17 -march=native -O3 -I../../include -I./ -I./dfscode -c main.cc
g++ -Wall -fopenmp -std=c++17 -march=native -O3 -I../../include -I./ -I./dfscode -c ../common/VertexSet.cc
g++ -Wall -fopenmp -std=c++17 -march=native -O3 -I../../include -I./ -I./dfscode -c ../common/graph.cc
g++ -Wall -fopenmp -std=c++17 -march=native -O3 -I../../include -I./ -I./dfscode main.o VertexSet.o graph.o omp_base.o -o fsm_omp_base -lgomp
mv fsm_omp_base ../../bin/
nvcc -gencode arch=compute_70,code=sm_70 -Xptxas -v -DUSE_GPU -O3 -w -DEDGE_PAR -I../../include -I./ -I./dfscode -c gpu_base.cu
/usr/local/cuda-12.2//bin/../targets/x86_64-linux/include/cub/agent/agent_batch_memcpy.cuh(886): error: expected an identifier
constexpr uint32_t (256 / 32) = BLOCK_THREADS / (1 << (5));
^

1 error detected in the compilation of "gpu_base.cu".
make: *** [../common.mk:123: gpu_base.o] Error 2

Error seems to be due to marco WARPS_PER_BLOCK (defined in common.h:59) which is also used as variable in name in agent_batch_memcpy.cuh:864

Hello! I found a small oversight in your code, and I hope you can fix it.
position: src\fsm\canonical.h
function name: is_frequent()
problem: I found that when it judges whether the pattern is frequent, it actually does not make a logical judgment, but only judges the size of pattern.

When I want to run FSM code on dataset mico, a segment fault occured. I think it may be caused by the incorrect usage of graph construction method in src/fsm/main.cc:21. However, when I fix that error by adding an argument to the call(changed to Graph g(argv[1], 0, 1, 1); ) , it gives me a error 700: Cuda error in file 'gpu_base.cu' in line 493 : an illegal memory access was encountered.

The command that I used:

./bin/fsm_gpu_base ./inputs/mico/graph 2 300

Could you please help me to fix that error?

Hello, there may be a lack of gpu_base.cu in src/fsm to build fsm_gpu_base as the readme did. Hope to solve the omission problem. Thank you.

In file fsm/gpu_base.cu, under the original setting of the MAX_NUM_PATTERNS, which means it equals to 8,000, the program gives an error code=700(cudaErrorIllegalAddress) "cudaDeviceSynchronize()", but if scaling it up to 13,000 or larger, the program runs successfully. So I have a question about how to set the value of this variable correctly, and if I don't do so, I have trouble debugging to get the program to run properly. Look forward for your help and thank you very much !

After running make in /src/pangolin/fsm, encounter an error when running ./bin/pangolin/./fsm_base ./inputs/citeseer/graph 3 100 as the tips.
The running details are as follows:

max_size = 3
min_support = 100
Number of unique labels: 6
|V| 3312 |E| 9072
Found 1 devices
Device[0]: NVIDIA GeForce GTX 1080 Ti
Compute capability: 6.1
Warp size: 32
Total # SM: 28
Total # CUDA cores: 3584
Total amount of shared memory per block: 49152 bytes
Total # registers per block: 65536
Total amount of constant memory: 65536 bytes
Total global memory: 10.9 GB
Memory Clock Rate: 5.25 GHz
Memory Bus Width: 352 bits
Peak Memory Bandwidth: 484.44 GB/s

error 1: Cuda error in file '../../../include/pangolin_gpu/../graph_gpu.h' in line 108 : invalid argument.

which corresponds to CUDA_SAFE_CALL(cudaMemcpy(d_labels_frequency, hg.get_label_freq_ptr(), (num_vertex_classes+1) * sizeof(vidType), cudaMemcpyHostToDevice));
after reading the source code, I found that the use of hg.get_label_freq_ptr() after labels_frequency_.clear(); may be the cause of the problem. But I have no idea how to modify it to make the program run properly.
Could you help me to solve this problem?
Best regards.