# Performance Study of GPUDirect Storage for CMS Reconstruction Workflows

- Institution: CERN
- Speaker: Ayaan Aziz Dewani
- Year: 2026
- Language: English
- Source: [CERN](https://videos.cern.ch/record/3026685)
- Licence: [CERN](https://copyright.web.cern)
- Who can watch: Teens profiles and older (sources give no ratings; placed by Invisible Network from the source's own words)

**Project Description**: Modern high-energy physics experiments such as CMS increasingly rely on heterogeneous computing architectures combining CPUs, GPUs, and high-performance storage systems. While GPU acceleration has significantly improved compute throughput, I/O and data movement between storage, host memory, and device memory remain a critical bottleneck. GPUDirect Storage (GDS) is a recent technology that allows GPUs to directly access data from NVMe storage, bypassing the CPU and reducing latency and memory copies. The goal of this project is to study, benchmark, and optimize the use of GPUDirect Storage and to evaluate how GDS impacts throughput and latency compared to traditional I/O paths using host memory.

**Supervisor**: Markus Holzer, Felice Pantaleo

00:00:00 Slide 1

00:00:14 Slide 2

00:01:20 Slide 3

00:02:18 Slide 4

00:03:13 Slide 5

00:04:35 Slide 6

00:05:16 Slide 7

00:06:38 Slide 8

00:07:07 Slide 9

00:08:03 Slide 10

00:08:44 Slide 11

00:09:02 Slide 12

00:11:00 Slide 13

00:11:44 Slide 14

## Lectures

1. [Performance Study of GPUDirect Storage for CMS Reconstruction Workflows](https://invisiblenetwork.ai/learn/cern/3026685/d6d4cf2b621b), 12 min (CERN's player)