The GFE (Graph Framework Evaluation) Driver is the program used to run the experiments in "BACH: Bridging Adjacency List and CSR Format using LSM-Trees for HGTAP Workloads", measuring the throughput of updates in libraries supporting structural dynamic graphs and the completion times of the Graphalytics kernels. The driver supports the following systems: Sortledton, Teseo, LLAMA, GraphOne, Stinger, LiveGraph, BACH and RocksDB. Additionally, it supports running the microbenchmarks from the Sortledton paper: Microbenchmarks. It can run three kinds experiments: insert all edges in a random permuted order from an input graph, execute the updates specified by a graphlog file and run the kernels of the Graphalytics suite: BFS, PageRank (PR), local triangle counting (LCC), weighted shortest paths (SSSP), weakly connected components (WCC) and community detection through label propagation (CDLP).

• O.S. Linux
• Autotools, Autoconf 2.69+
• A C++17 compliant compiler with support for OpenMP. We tested it with GCC 10.
• libnuma 2.0 +
• libpapi 5.5 +
• SQLite 3.27 +
• Intel Threading Building Blocks 2 (version 2020.1-2)
• Disable NUMA balancing feature to avoid the Linux Kernel to swap pages during insertions: echo 0 | sudo tee /proc/sys/kernel/numa_balancing

Initialise the sources and the configure script by:

git clone https://github.com/PerFuchs/gfe_driver
cd gfe_driver
git submodule update --init
mkdir build && cd build
autoreconf -iv ..

The driver needs to be linked with the system to evaluate, which has to be built ahead. We do not recommend linking the driver with multiple systems at once, due to the usage of global variables in some systems and other naming clashes. Instead, it is safer to reconfigure and rebuild the driver each time for a single specific system.

Download the library from https://github.com/2600254/BACH_demo.

mkdir build && cd build
cmake ..
make

Then configure the driver by pointing the path to the compiled dynamic library is located:

mkdir build && cd build
../configure --enable-optimize --disable-debug --with-bach=/path/to/BACH_demo/build

apt-get install librocksdb-dev
git clone https://github.com/facebook/rocksdb

Then configure the driver by pointing the path to the header files of RocksDB:

mkdir build && cd build
../configure --enable-optimize --disable-debug --with-rocksdb=/path/to/rocksdb

Use the branch feature/gfe , it contains additional patches w.r.t. upstream, from https://github.com/whatsthecraic/stinger. For the paper, we evaluted commit "2bcfac38785081c7140b0cd27f3aecace088d664"

git clone https://github.com/whatsthecraic/stinger -b feature/gfe
cd stinger
mkdir build && cd stinger
cmake ../ -DCMAKE_BUILD_TYPE=Release -DBUILD_TESTING=0 
make

If the build has been successful, it should at least create the executable bin/stinger_server.

Configure the GFE driver with:

mkdir build && cd build
../configure --enable-optimize --disable-debug --with-stinger=/path/to/stinger/build

Use the branch feature/gfe , it contains additional patches w.r.t. upstream, from https://github.com/whatsthecraic/llama. For the paper, we evaluated commit 32053f4ff800bed1989da79e189feeaa1bbb84b3.

git clone https://github.com/whatsthecraic/llama -b feature/gfe

LLAMA is a header-only library. It does not need to be compiled in advance.

Configure the GFE driver with:

mkdir build && cd build
../configure --enable-optimize --disable-debug --with-llama=/path/to/llama

Use the branch feature/gfe , it contains additional patches w.r.t. upstream, from https://github.com/whatsthecraic/GraphOne. For the paper, we evaluated "1475bf5887aaf37dd7aa47377e9f11a94aa0d880".

git clone https://github.com/whatsthecraic/GraphOne -b feature/gfe
cd GraphOne
mkdir build && cd build
cmake -S ../ -DCMAKE_BUILD_TYPE=Release
make -j

If the build has been successful, it should at least create the executable graphone64. Then, configure the driver with:

mkdir build && cd build
../configure --enable-optimize --disable-debug --with-graphone=/path/to/graphone/build

Download the binary library from the official repository. In the paper, we evaluated version 20200829. Then configure the driver by pointing the path to where the library has been downloading:

mkdir build && cd build
../configure --enable-optimize --disable-debug --with-livegraph=/path/to/livegraph/lib

Use the branch master from https://github.com/cwida/teseo. In the paper, we evaluated version 14227577731d6369b5366613f3e4a679b1fd7694.

git clone https://github.com/cwida/teseo
cd teseo
./autoreconf -iv
mkdir build && cd build
../configure --enable-optimize --disable-debug
make -j

If the build has been successful, it should at least create the archive libteseo.a. Then configure the driver with:

mkdir build && cd build
../configure --enable-optimize --disable-debug --with-teseo=/path/to/teseo/build   

Use the branch master from https://gitlab.db.in.tum.de/per.fuchs/sortledton. For the paper, we evaluated commit "a32b8ac208bb889b518e14b1317957c9a8c466b6".

Follow the instructions in the README of the repository to setup and build the library. Then configure the driver with:

mkdir build && cd build
../configure --enable-optimize --disable-debug --with-sortledton=/path/to/microbenchmark/build   

Use the branch master from https://gitlab.db.in.tum.de/per.fuchs/graph-data-structure-microbenchmarks. For the paper, we evaluated commit "a32b8ac208bb889b518e14b1317957c9a8c466b6".

Follow the instructions in the README of the repository to setup and build the library. Then configure the driver with:

mkdir build && cd build
../configure --enable-optimize --disable-debug --with-microbenchmarks=/path/to/microbenchmark/build   

Once configured, run make -j. There is no install target, the final artifact is the executable gfe_driver.

If in the mood of running the testsuite, type make check -j.

In our experiments, we used the following input graphs and data sets:

• dota-league and graph500-SF, with SF in {22, 24 26} and com-friendster, were taken from the official Graphalytics collection.
• uniform-SF, with SF in {22, 24, 26} were generated with an ad-hoc tool. These are synthetic graphs having the same number of vertices and edges of graph500-SF, but a uniform node degree distribution.
• The logs for the experiments with updates, i.e. with both insertions and deletions, were generated with another ad-hoc tool.
• yahoo-songs and edit-enwiki were taken from the Konect webpage they were prepared for our experiments by sorting them by timestamp and removing duplicates by using tools/timestampd_graph_2_edge_list.py.
• uk-2007-05 can be found at there, which is derived by reformatting the data from there.

A complete image of all datasets used in the experiments can be downloaded from Zenodo: input graphs, graph logs, dense friendster and timestamped graphs.

The driver takes as input a list of options together with a graph, and emits the results into a sqlite3 database. There are three kinds of experiments that can be executed:

• Insertions only: insert all vertices and edges from an input graph, in a random order. Use the command:

./gfe_driver -G /path/to/input/graph.properties -u -l <system_to_evaluate> -w <num_threads> -d output_results.sqlite3

For LLAMA only: add the option --build_frequency 10s to asynchronously issue the creation of a new level (or delta) every 10 seconds.

• Updates: perform all insertions and deletions from a log. Add the option --log /path/to/updates.graphlog :

./gfe_driver -G /path/to/input/graph.properties -u --log /path/to/updates.graphlog --aging_timeout 24h -l <system_to_evaluate> -w <num_threads> -d output_results.sqlite3

• Graphalytics: execute the six kernels from the Graphalytics suite. Add the option -R <N> to repeat N times the execution of all Graphalytics kernels, one after the other. E.g., to run the kernels five times, after all vertices and edges have been inserted, use:

./gfe_driver -G /path/to/input/graph.properties -u -l <system_to_evaluate> -w <num_threads> -R 5 -d output_results.sqlite3

Type ./gfe_driver -h for the full list of options and for the libraries that can be evaluated (option -l). The driver spawns the number of threads given by the option -w to concurrently run all insertions or updates. For Graphalytics, it defaults to the total number of the physical threads in the machine. This setting can be changed with the option -r <num_threads>. Note that the numbers in the library codes (e.g. teseo.6, stinger3) are unrelated to the versions of the systems evaluated, they were only used internally for development purposes.

The database output_results.sqlite3 will contain the final results. Refer to this repository to see how to load and inspect the data within Jupyter notebooks and how to recreate the same plots of the paper.

These are the full commands to repeat the experiments in the paper:

libraries=("bach_ro_leveling" "bach_ro_tiering" "livegraph3_ro" "rocksdb" "llama8-ref")
datasets=("datagen-8_5-fb" "graph500-26" "com-friendster" "uk-2007-05")
for dataset in "${datasets[@]}"; do
  for library in "${libraries[@]}"; do
    ./gfe_driver -u -d results.sqlite3 -l $library -G /path/to/input/$dataset.properties -w 16 --latency true 
    ./gfe_driver -u -d results.sqlite3 -l $library -G /path/to/input/$dataset.properties -w 16 --latency true --is_timestamped true
    ./gfe_driver -u -d results.sqlite3 -l $library -G /path/to/input/$dataset.properties -w 16 --log /path/to/input/$dataset-0.3.graphlog --aging_timeout 10h --latency true --is_timestamped true
  done
done

libraries=("bach_ro_leveling" "bach_ro_tiering" "livegraph3_ro" "rocksdb" "llama8-ref")
datasets=("datagen-8_5-fb" "graph500-26" "com-friendster" "uk-2007-05")
for dataset in "${datasets[@]}"; do
  for library in "${libraries[@]}"; do
    ./gfe_driver -u -R 5 -d results.sqlite3 -l $library -G /path/to/input/$dataset.properties -w 16  
  done
done

For all graphs with 5 runs.

libraries_bach=("bach_ro_elastic" "bach_ro_tiering" "bach_ro_leveling")
datasets=("datagen-8_5-fb" "graph500-26" "com-friendster")
for dataset in "${datasets[@]}"; do
  for library in "${libraries_bach[@]}"; do
    ./gfe_driver -u -d results.sqlite3 -l $library -G /path/to/input/$dataset.properties -w 16 -r 16 --log /path/to/input/$dataset-0.5.graphlog --aging_timeout 10h --mixed_workload true -R 1
  done
done
libraries=("livegraph3_ro" "rocksdb")
for dataset in "${datasets[@]}"; do
  for library in "${libraries[@]}"; do
    ./gfe_driver -u -d results.sqlite3 -l $library -G /path/to/input/$dataset.properties -w 16 -r 16 --log /path/to/input/$dataset-0.5.graphlog --aging_timeout 10h --mixed_workload true -R 1
  done
done
for dataset in "${datasets[@]}"; do
  for library in "${libraries_bach[@]}"; do
    ./gfe_driver -u -d results.sqlite3 -l $library -G /path/to/input/graph500-26.properties -w 16 -r 16 --log /path/to/input/graph500-26-0.5-sq.graphlog --aging_timeout 10h --mixed_workload true -R 1 --is_timestamped true
  done
done
for dataset in "${datasets[@]}"; do
  for library in "${libraries[@]}"; do
    ./gfe_driver -u -d results.sqlite3 -l $library -G /path/to/input/graph500-26.properties -w 16 -r 16 --log /path/to/input/graph500-26-0.5-sq.graphlog --aging_timeout 10h --mixed_workload true -R 1 --is_timestamped true
  done
done

The option --aging_timeout serves to limit the total time to execute the experiment. For LLAMA, it could be necessary to stop the experiment earlier, as the continuous creation of new deltas can cause a memory exhaustion.
For the experiment with the memory footprint of Figure 7d, add the arguments: --aging_memfp --aging_memfp_physical --aging_memfp_threshold 330G --aging_release_memory=false. The option --aging_memfp records the memory footprint as the experiment proceeds, --aging_memfp_physical records the physical memory (RSS) of the process, rather than the virtual memory of the glibc allocator, --aging_memfp_threshold 330G terminates the experiment if the memory footprint measured is greater than 330 GB and --aging_release_memory=false avoids releasing the memory used in the driver to load the graph from the file, as it may (or may not) recycled by the libraries. With the memory footprint, for LLAMA, it"s not necessary to set --aging_timeout 4h as --aging_memfp_threshold 330G already acts as a guard on the overall memory consumption.