Add latest cuGraph

[gitbuda]

• [ ] Follow up HERE (Discord)
• [ ] Make sure all API changes produce correct results -> TESTING 🧪
  • [ ] cugraph.balanced_cut_clustering
  • [ ] cugraph.betweenness_centrality
  • [ ] cugraph.generator
  • [ ] cugraph.hits
  • [ ] cugraph.katz_centrality
  • [ ] cugraph.leiden ⏳
  • [ ] cugraph.louvain
  • [x] cugraph.pagerank
  • [ ] cugraph.personalized_pagerank
  • [ ] cugraph.spectral_clustering
• [ ] Figure out undefined symbols under laiden and betweenness_centrality
• [ ] Upgrade Python pip libs for the full mage with cugraph Docker image
• [ ] Consider adding a template generator for Dockerfiles or some form of includes -> https://codeberg.org/devthefuture/dockerfile-x#include
• [ ] Upgrade dgl -> any build (native/Docker) fails -> figure out how to compile
  • [x] Try with https://catalog.ngc.nvidia.com/orgs/nvidia/containers/dgl
  • [x] Install nvidia-container-toolkit; restart Docker systemctl restart docker; run docker run -it --rm --gpus all ubuntu nvidia-smi to verify (source)
  • [ ] Both cuGraph and DGL projects recommend using conda -> use conda under the container
• [ ] Extend mage/setup to pick individual libraries to install -> add that as a docker argument because usually, someone needs a specific binary (while the whole think doesn't compile); consider adding init bash script because usually, installation of deps fails (e.g. gdl is a very complex dependency)

Docker Build Commands

docker build -f Dockerfile.experiment_partial -t mage-cugraph-part --progress plain .
docker run -it --rm --gpus all mage-cugraph-part bash

docker build -f Dockerfile.experiment_full -t mage-cugraph-full --progress plain .

docker run -it --rm --gpus all mage-cugraph-full bash
cd /mage/cpp/build
cmake -DMAGE_CUGRAPH_ENABLE=ON -DMAGE_CUGRAPH_ROOT=/opt/conda ..
VERBOSE=1 make cugraph.pagerank
make cugraph.pagerank cugraph.personalized_pagerank cugraph.louvain cugraph.katz_centrality cugraph.leiden cugraph.betweenness_centrality cugraph.balanced_cut_clustering cugraph.spectral_clustering cugraph.hits cugraph.generator

python3 /mage/setup build --gpu --cpp-build-flags CMAKE_BUILD_TYPE=Release MAGE_CUGRAPH_ROOT=/opt/conda/ -p /usr/lib/memgraph/query_modules/

docker run -it --rm --gpus all -p 7687:7687 mage-cugraph-full --log-level=TRACE --also-log-to-stderr

# if building memgraph, do it outside the /mage/cpp/memgraph because that's copied to during Docker build
cd /tmp/mage/cpp/dist
cp ../build/cugraph_module/cugraph.pagerank.so ./

cd /tmp/mage && ./test_e2e -k "pagerank_test-test_cugraph_influential"

Docker Base Images

• https://hub.docker.com/r/rapidsai/base/tags
• https://hub.docker.com/r/nvidia/cuda/tags
• https://nvidia.github.io/cccl/libcudacxx/

Done

• [x] Compile only the cuGraph module with latest everything
  • [x] python3 /mage/setup build --gpu --cpp-build-flags MAGE_CUGRAPH_ROOT=/opt/conda CMAKE_C_COMPILER=gcc CMAKE_CXX_COMPILER=g++ -p /usr/lib/memgraph/query_modules
  • [x] Figure out what's wrong with fmt (when all is compiled), again issues with fmt+spdlog, spdlog from /opt/conda is not ok (maybe interesting https://github.com/gabime/spdlog/issues/2825); fmt v9 doesn't compile with used g++, fmt v10 doesn't work with the spdlog in the /opt/conda
    • [x] Potential solution -> https://github.com/gabime/spdlog/issues/1897
  • [x] cugraph/algorithms.hpp for some reason doesn't work -> make sure GPU+Docker is ok
  • [x] Figure out the problem with cuda code under conda -> most likely some build config issue 🤔 -> https://github.com/NVIDIA/cuCollections/issues/331#issuecomment-1866958019
  • [x] API fixes
• [x] Find and link libcugraph-ops++.so and libraft.so -> taken from rapidsai/base Docker image under /opt/conda/lib
• [x] Make cugraph.pagerank to work correctly, implement the package process and release Docker image only with memgraph and pagerank internally

Experiment / Optional

• [ ] Take a look at the edge mask feature (there is no masking of vertices)

---

Description

Please briefly explain the changes you made here.

Pull request type

• [ ] Bugfix
• [ ] Algorithm/Module
• [ ] Feature
• [ ] Code style update (formatting, renaming)
• [ ] Refactoring (no functional changes, no api changes)
• [ ] Build related changes
• [ ] Documentation content changes
• [ ] Other (please describe):

Related issues

Delete if this PR doesn't resolve any issues. Link the issue if it does.

######################################

Reviewer checklist (the reviewer checks this part)

Module/Algorithm

• [ ] Core algorithm/module implementation
• [ ] Query module implementation
• [ ] Tests provided (unit / e2e)
• [ ] Code documentation
• [ ] README short description

Documentation checklist

• [ ] Add the documentation label tag
• [ ] Add the bug / feature label tag
• [ ] Add the milestone for which this feature is intended
  • If not known, set for a later milestone
• [ ] Write a release note, including added/changed clauses
  • [Release note text]
• [ ] Link the documentation PR here
  • [Documentation PR link]
• [ ] Tag someone from docs team in the comments

[nad010286]

any update? look like cugraph.pagerank still results different output compared with pagerank cpu version

[gitbuda]

@nad010286 nothing yet, just didn't have a chance to finish this work, but it's still on the short-term TODO list 😄

[gitbuda]

Hi @nad010286! I've made some progress on a small scale. Seems like the regular CPU pagerank and cugraph pagerank are producing the same outputs (CPU test case vs cugraph test case) 🤔

Can you please provide a test case from your side (something reasonable in size, max 100k nodes + edges), in the e2e test format 🙏:

• The graph should be in the .cyp format (it's actually cypher per line, it can be many lines, but it's important to not break down single query into multiple lines) GRAPH INPUT EXAMPLE
• The test case should be in the .yml format (make sure you have everything, the exact query, and exact outputs) TEST CASE EXAMPLE

[sonarcloud[bot]]

Quality Gate failed

Failed conditions
6 Security Hotspots

See analysis details on SonarCloud

[nad010286]

Hi @nad010286! I've made some progress on a small scale. Seems like the regular CPU pagerank and cugraph pagerank are producing the same outputs (CPU test case vs cugraph test case) 🤔

Can you please provide a test case from your side (something reasonable in size, max 100k nodes + edges), in the e2e test format 🙏:

• The graph should be in the .cyp format (it's actually cypher per line, it can be many lines, but it's important to not break down single query into multiple lines) GRAPH INPUT EXAMPLE
• The test case should be in the .yml format (make sure you have everything, the exact query, and exact outputs) TEST CASE EXAMPLE

hey, Im having issue to build the cuGraph and MAGE from source :( any instructions or things I have to pay attention to? Or if you can send me the docker image that you already built, Im happy to perform some tests for you with loads of graph data

[intuitiveminds]

Hi @nad010286! I've made some progress on a small scale. Seems like the regular CPU pagerank and cugraph pagerank are producing the same outputs (CPU test case vs cugraph test case) 🤔 Can you please provide a test case from your side (something reasonable in size, max 100k nodes + edges), in the e2e test format 🙏:

• The graph should be in the .cyp format (it's actually cypher per line, it can be many lines, but it's important to not break down single query into multiple lines) GRAPH INPUT EXAMPLE
• The test case should be in the .yml format (make sure you have everything, the exact query, and exact outputs) TEST CASE EXAMPLE

hey, Im having issue to build the cuGraph and MAGE from source :( any instructions or things I have to pay attention to? Or if you can send me the docker image that you already built, Im happy to perform some tests for you with loads of graph data

I would also be very interested in an updated guide on how to build cugraph mage - I tried building the main branch with CUDA 11.8 but the nvcc compiler does not support C++ 20, so it fails with unorder_map::contains later on.

[nad010286]

managed to build and run some tests @gitbuda The results are similar with very small difference. But processing time is worse than CPU. Not sure if GPU is used because by monitoring the GPU via nvidia-smi, I can only see power consumption increase after the result is out. Also no VRAM is used.