Computational Pangenoemics is an emerging field that studies genetic variation using a graph structure encom- passing multiple genomes. Visualizing pangenome graphs is vital for understanding genome diversity. Yet, handling large graphs can be challenging due to the high computational demands of the graph layout process.
Evaluated on 24 human whole-chromosome pangenomes, our GPU-based solution achieves a 57.3x speedup over the state-of-the-art multithreaded CPU baseline odgi-layout without layout quality loss, reducing execution time from hours to minutes.
We build upon the widely-used pangenome tools ODGI. This codebase includes our GPU kernel implementation of odgi-layout.
Update: now the GPU-supported build is also integrated into the ODGI. You can follow the ODGI's installation guide to try GPU-accelerated layout easily!
For CPU-only build, you can follow the installation guide in ODGI.
Our requirements are the same:
odgi requires a C++ version of 9.3 or higher.
odgi pulls in a host of source repositories as dependencies. It may be necessary to install several system-level libraries to build odgi. On Ubuntu 20.04, these can be installed using apt:
sudo apt install build-essential cmake python3-distutils python3-dev libjemalloc-dev
To build with GPU, we provide a flag USE_GPU when cmake.
git clone --recursive git@github.com:tonyjie/odgi.git
cd odgi
mkdir build
cmake .. -DCMAKE_BUILD_TYPE=Generic -DUSE_GPU=ON
make -j
The runnable odgi binary should be at odgi/bin/odgi
To compute odgi layout with GPU is very simple: you only need an extra --gpu. The other arguments follows the description here.
An example is as follows:
odgi/bin/odgi layout -i ${INPUT_OG} -o ${OUTPUT_LAY} --threads ${NUM_THREAD} --gpu
Machines:
- 32-thread CPU baseline: 32-core Intel Xeon Gold 6246R CPU@3.4GHz.
- A6000: NVIDIA RTX A6000 GPU.
- A100: NVIDIA A100 GPU.
Run time and speedup — the run time format is in h:mm:ss.
| Chromosome | CPU | A6000 | Speedup (A6000) | A100 | Speedup (A100) |
|---|---|---|---|---|---|
| Chr.1 | 2:32:38 | 0:04:59 | 30.6x | 0:02:42 | 56.5x |
| Chr.2 | 1:17:03 | 0:03:33 | 21.7x | 0:01:01 | 75.8x |
| Chr.3 | 1:28:41 | 0:03:27 | 25.7x | 0:01:31 | 58.5x |
| Chr.4 | 1:47:32 | 0:03:40 | 29.3x | 0:02:06 | 51.2x |
| Chr.5 | 1:41:09 | 0:03:19 | 30.5x | 0:01:07 | 90.6x |
| Chr.6 | 1:13:55 | 0:02:49 | 26.3x | 0:01:27 | 51.0x |
| Chr.7 | 1:16:46 | 0:03:00 | 25.6x | 0:01:34 | 49.0x |
| Chr.8 | 1:17:27 | 0:02:57 | 26.3x | 0:01:41 | 46.0x |
| Chr.9 | 1:16:49 | 0:02:53 | 26.6x | 0:00:55 | 83.8x |
| Chr.10 | 0:48:34 | 0:02:22 | 20.6x | 0:00:44 | 66.2x |
| Chr.11 | 0:56:25 | 0:02:07 | 26.7x | 0:00:37 | 91.5x |
| Chr.12 | 0:44:05 | 0:02:07 | 20.9x | 0:00:49 | 54.0x |
| Chr.13 | 1:03:32 | 0:02:22 | 26.8x | 0:00:53 | 71.9x |
| Chr.14 | 0:51:21 | 0:02:04 | 24.9x | 0:00:46 | 67.0x |
| Chr.15 | 1:11:33 | 0:02:52 | 25.0x | 0:01:16 | 56.5x |
| Chr.16 | 2:19:47 | 0:04:56 | 28.3x | 0:12:58 | 10.8x |
| Chr.17 | 1:03:45 | 0:02:01 | 31.7x | 0:01:07 | 57.1x |
| Chr.18 | 0:50:29 | 0:01:50 | 27.6x | 0:01:08 | 44.6x |
| Chr.19 | 0:40:23 | 0:01:29 | 27.3x | 0:00:27 | 89.8x |
| Chr.20 | 0:51:34 | 0:01:30 | 34.3x | 0:01:01 | 50.7x |
| Chr.21 | 0:44:18 | 0:01:26 | 30.9x | 0:00:38 | 69.9x |
| Chr.22 | 0:39:59 | 0:01:37 | 24.8x | 0:00:30 | 80.0x |
| Chr.X | 1:04:06 | 0:01:49 | 35.4x | 0:00:49 | 78.4x |
| Chr.Y | 0:01:55 | 0:00:03 | 36.9x | 0:00:04 | 28.7x |
| Geometric Mean | 27.7x | 57.3x |
Ready to be merged into the ODGI with this PR.
To reproduce the experiments in the paper, the easiet way is to follow our provided artifact repo.
Rapid GPU-Based Pangenome Graph Layout: https://ieeexplore.ieee.org/abstract/document/10793207
Presented at SC'24.
@inproceedings{li2024rapid,
title={Rapid GPU-Based Pangenome Graph Layout},
author={Li, Jiajie and Schmelzle, Jan-Niklas and Du, Yixiao and Heumos, Simon and Guarracino, Andrea and Guidi, Giulia and Prins, Pjotr and Garrison, Erik and Zhang, Zhiru},
booktitle={SC24: International Conference for High Performance Computing, Networking, Storage and Analysis},
pages={1--19},
year={2024},
organization={IEEE}
}