Skip to content

Latest commit

 

History

History
111 lines (85 loc) · 4.18 KB

README.md

File metadata and controls

111 lines (85 loc) · 4.18 KB

Experiments

Table of Contents

Overview

Our experiments have been automated by scripts (run.py). Each figure or table in our paper is treated as one experiment and is associated with a subdirectory in fgnn-artifacts/exp. The script will automatically run the experiment, save the logs into files, and parse the output data from the files.

> tree -L 2 fgnn-artifacts/exp
fgnn-artifacts/exp
├── fig4a
├── fig4b
├── ...
├── table1
│   ├── README.md
│   ├── run.py
├── ...
├── Makefile

Paper's Hardware Configurations

  • 8 * NVIDIA V100 GPUs (16GB of memory each)
  • 2 * Intel Xeon Platinum 8163 CPUs (24 cores each)
  • 512GB RAM

Note: If you have a different hardware environment, you need to goto the subdirectories (i.e., figXX or tableXX), follow the instructions to modify some script configurations(e.g. smaller cache ratio), and then run the experiment

Run A Single Experiment

The following commands are used to run a certain experiment(e.g. table1).

cd fgnn-artifacts/exp
make table1.run

Moreover, the user can also goto a subdirectory (e.g., fgnn-artifacts/exp/table1) and then follow the instruction (README.md) to run the experiment.

Run All Experiments

The following commands are used to run all experiments. Note that running all experiments may take several hours. This table lists the expected running time for each experiment.

cd fgnn-artifacts/exp
make all

Example output.

The experiment output files are in the subdirectories (figXX/run-logs or figXX/output_XX). The output files include log files for each testcase, parsed data, and eps-format figures.

> cat output_2022-01-29_17-19-44/table1.dat
GNN Systems               Sample  Extract  Train  Total    #
DGL                         4.40    13.63   4.10  22.51    # logs_dgl/test1.log
 w/ GPU-base Sampling       1.28    13.61   4.12  19.08    # logs_dgl/test0.log
SGNN                        3.26     5.96   4.14  13.37    # logs_sgnn/test3.log
 w/ GPU-base Caching        3.05     1.95   4.07   8.99    # logs_sgnn/test2.log
 w/ GPU-base Sampling       0.72     5.88   4.07  10.70    # logs_sgnn/test1.log
 w/ Both                    0.72     3.90   3.98   8.61    # logs_sgnn/test0.log

Expected Running Time

Experiment Number of Test Cases Expected Run Time
Fig4a 31 tests 40 mins
Fig4b 31 tests 40 mins
Fig5a 31 tests 40 mins
Fig5b 36 tests 40 mins
Fig10 48 tests 50 mins
Fig11a 73 tests 80 mins
Fig11b 94 tests 120 mins
Fig11c 94 tests 120 mins
Fig12 33 tests 30 mins
Fig13 33 tests 30 mins
Fig14a 34 tests 32 mins
Fig14b 34 tests 23 mins
Fig15 36 tests 40 mins
Fig16a 3 tests 25 mins
Fig17a 14 tests 20 mins
Fig17b 27 tests 28 mins
Table1 12 tests 10 mins
Table2 12 tests 15 mins
Table4 44 tests 45 mins
Table5 32 tests 35 mins
Table6 4 tests 5 mins

Clean Experiment Logs

make clean

FAQ

The data mismatch the paper data.

  • Do not run nvidia-smi command during the tests. nvidia-smi command harm the performance severely.