IA3_2022

cuDF vs Pandas DataFrame performance comparison

The following sections include the system configuration, the dataset, the flowchart, and the benchmarking result.

System configuration

• GPU information:
  • NVIDIA A100-SXM4-40GB
  • Driver Version: 470.129.06
  • CUDA Version: 11.4
• Hardware information:
  • Total Memory: 1.0T
  • CPU name: AMD EPYC 7742 64-Core Processor
  • CPU(s): 256
• OS information:
  • Operating System: Ubuntu 20.04.4 LTS
  • Kernel: Linux 5.4.0-121-generic
  • Architecture: x86-64
• Python package information:
  • Python version: 3.9.13
  • Conda version: conda 4.13.0
  • cuda-python: 11.7.0
  • cudatoolkit: 11.2.72
  • cudf: 22.06.01
  • pandas: 1.4.3

Dataset

We used datasets from the Stanford Large Network Dataset Collection and the Real Datasets for Spatial Databases: Road Networks and Points of Interest.

Graph	Type	V	E
p2p-Gnutella09	Directed	8,114	26,013
p2p-Gnutella04	Directed	10,876	39,994
Skitter	Undirected	1,696,415	11,095,298
roadNet-CA	Undirected	1,965,206	2,766,607
roadNet-TX	Undirected	1,379,917	1,921,660
roadNet-PA	Undirected	1,088,092	1,541,898
SF.cedge	Undirected	1,74,955	2,23,001
cal.cedge	Undirected	21,048	21,693
TG.cedge	Undirected	18,263	23,874
OL.cedge	Undirected	6,105	7,035

• V represents the number of nodes and E represents the number of edges of the graph.

Benchmarks

We calculate execution times to calculate the basic relational operations, triangle counting, and transitive closure computation for both cuDF and Pandas DF on the same dataset in the mentioned environment:

• To generate result for basic operations using cuDF and Pandas:

python basic_operations.py

• To generate result for triangle counting using Rapids cuDF:

python triangle_counting.py

• To generate result for triangle counting using Pandas DF:

python triangle_counting_pandas.py

• To generate result for transitive closure using Rapids cuDF:

python transitive_closure.py

• To generate result for transitive closure using Pandas DF:

python transitive_closure_pandas.py

Flowchart (Transitive closure)

flowchart TD
    B["result = G"] --> C["r = #961;<sub>0/1</sub>(G)"]
    C --> D["projection = #960;<sub>1, 2</sub>(r #8904;<sub>0</sub> G)"]
    D --> E["old_result_size = length(result)"]
    E --> F["result = projection #8899; G"]
    F --> G["new_result_size = length(result)"]
    G --> H{Is old_result_size = new_result_size?}
    H --> |Yes| I[return result]
    H --> |No| J["r = #961;<sub>0/1</sub>(projection)"]
    J --> D

Loading

Pandas vs CUDF performance comparison

• Benchmark for triangle counting

Dataset	Triangles	cuDF(s)	Pandas(s)	Speedup
Skitter	28,769,868	1.0340	152.7032	147.7x
roadNet-CA	120,676	0.1593	3.0106	18.9x
roadNet-TX	82,869	0.1180	1.9466	16.5x
roadNet-PA	67,150	0.1084	1.4777	13.6x
SF.edge	4,036	0.0877	0.1482	1.7x
p2p-Gnutella09	2,354	0.0486	0.0256	0.5x

• Benchmark for transitive closure

Dataset	TC size	Iterations	cuDF(s)	Pandas(s)	Speedup
SF.cedge	80,498,014	287	64.235	4582.067	71.3x
p2p-Gnutella09	21,402,960	20	3.881	167.143	43.1x
p2p-Gnutella04	47,059,527	26	14.104	569.249	40.4x
cal.cedge	501,755	195	3.883	5.769	1.5x
TG.cedge	481,121	58	1.191	1.400	1.2x
OL.cedge	146,120	64	0.557	0.474	0.9x

Figures

• Figure: Comparison for basic operations

• Figure: Comparison for triangle counting

Figure: Comparison for transitive closure

Reference

• Documentation on CUDF Drop
• Documentation on CUDF Drop Duplicates
• Documentation on CUDF concatenate
• California road network
• (Leskovec 2009) J. Leskovec, K. Lang, A. Dasgupta, M. Mahoney. Community Structure in Large Networks: Natural Cluster Sizes and the Absence of Large Well-Defined Clusters. Internet Mathematics 6(1) 29--123, 2009.
• Real Datasets for Spatial Databases: Road Networks and Points of Interest
• Stanford Large Network Dataset Collection
• Flowcharts - Basic Syntax