desktop, hpc and cloud big data pipelines for gpu-accelerated...gpu-accelerated big data pipelines...
TRANSCRIPT
GPU-AcceleratedBig Data Pipelines for
Desktop, HPC and CloudDr. Melissa Smith1, Ben Shealy1, Josh Burns2, Dr. Alex Feltus3, Dr. Stephen Ficklin2
1 Department of Electrical and Computer Engineering, Clemson University2 Department of Horticulture, Washington State University
3 Department of Genetics and Biochemistry, Clemson University
2
Overview
- KINC
- KINC Nextflow Pipeline
- Running KINC Pipeline on Kubernetes
- Demo
- Challenges / Opportunities
3
The Gene Co-Expression Network (GCN)
4
Knowledge Independent Network Construction (KINC)
→
Gene expression matrix
Gene co-expression network
Similarity matrix
Pairwise scatter plot
↓
5
Human Brain Tissue-Specific Network
6
Human Kidney Tumor-Specific Network
7
KINC GPU Implementation
Shealy, Burns, et. al., “GPU Implementation of Pairwise Gaussian Mixture Models for Multi-Modal Gene Co-Expression Networks”, IEEE Access
8
KINC Pipeline
9
Pipeline Portability with Nextflow
# Local
nextflow run systemsgenetics/KINC-nf -with-docker
# HPC
nextflow run systemsgenetics/KINC-nf -profile pbs -with-singularity
# Kubernetes
nextflow kuberun systemsgenetics/KINC-nf -v <pvc-name>
10
Running Genomics Pipelines on Kubernetes
11
Integrating Nextflow / Kubernetes / GPUs
nextflow.config
process {
cpus = 4
memory = 8.GB --->
accelerator = 1
}
pod.yaml
resources:
limits:
cpu: 4
memory: "8Gi"
nvidia.com/gpu: 1
12
DemoKINC / Nextflow / Kubernetes / NVIDIA GPUs
13
Challenges and Opportunities
- Using Kubernetes as an HPC platform- Transient failures (network timeouts, node-specific issues)
- Scheduling policies (walltime)
- Access controls (users)
- “SLURM-enetes” / “KubLURM”
- Scale up
- Increase usability for domain scientists- Friendly user interface
- Assist user with machine learning 14
Thank you! Questions?Research funded by NSF Grant #1659300
15
Nextflow API / Workflows
16
Nextflow API / Workflow Instance
17
Grafana / Cluster Usage
18
Grafana / GPU Usage
19