In collaboration with our partner NVIDIA, online introductory training was delivered on 5th and 6th November 2024 to prepare Hackathon participants as well as giving our users and members of the wider HPC community the chance to get insight of NVIDIA’s latest technologies.
This training was delivered in two online sessions and were joined by more than fourty participants.
The Grace Hopper Superchip combines the high performance of the Hopper GPU and the versatility of the Grace Arm CPU with a high-bandwidth NVLink C2C and cache-coherent memory access between the CPU and GPU. Details of Grace Hopper architecture, programming model and Unified Memory over NVLink C2C were explored. How existing applications benefit from Grace Hopper, and how Grace Hopper opens doors to bring new applications to accelerated computing platforms were also covered in the session.
Presenters:
Paul Graham – Senior Solutions Architect at NVIDIA
Filippo Spiga- Technical Product Manager, Accelerated Compute Workloads and Performance, NVIDIA
In this session participants were introduced two of the Nsight Developer Tools (https://developer.nvidia.com/tools-overview) used for performance analysis.
Nsight Systems is a performance analysis tool aimed at analysing an application at the system level. It is designed to visualise an application’s algorithms, identify the largest opportunities for optimisation, and tune to scale efficiently across any quantity or size of CPUs and GPUs, from large servers to our smallest systems-on-a-chip (SoCs). This talk comprised the tracing process, including project setup and configuration and how to specify profiling targets, launch an application to trace, and view the results. It covered how to read and analyse an Nsight Systems report, specifically the timeline view, which provides an intuitive visualisation of system events, making it easier to understand an application’s behaviour and the CPU-GPU interactions.
Nsight Compute allows for a deeper dive into the compute processes executing on the GPU. It visualises hardware metrics and presents them in an intuitive away along with advice and actionable information. In the latter part of the session, participants were given insights into the capabilities of Nsight Compute, including setup tips and key features for performance analysis and how to better understand and quantify the performance of their GPU. The following was also explored: how Nsight Compute collects performance and throughput metrics, including from hardware counters and code instrumentation, how this can be correlated with source code, and how the guided analysis this enables can be used to understand and optimise participants’ application.
Presenters:
– Robert Dietrich – Senior System Software Engineer at NVIDIA
– Felix Schmitt – Senior System Software Engineer at NVIDIA
From 26th to 28th November 2024, DIRAC, in collaboration with NVIDIA and the N8 Research Partnership, hosted a hackathon at Durham University focused on the porting and optimisation of DiRAC codes.
Nine teams, consisting of more than thirty participants, were supported by three NVIDIA experts, two DiRAC RSEs, and one N8 RSE throughout the event.
This three-day hackathon provided an excellent opportunity for participants to gain insights and hands-on guidance with NVIDIA’s latest technologies: the A100 and Grace Hopper systems.
To help participants make the most of the event, two online training sessions (insert link) were provided in advance, along with early access to resources. Onboarding sessions, organised with the systems administrators from Tursa and N8, ensured seamless access to both the A100 and Grace Hopper systems for all participants.
Before the event, NVIDIA experts also offered one-to-one sessions with each team, allowing participants and support staff the chance to pre-empt and address potential challenges.
To conclude the event, a cold buffet at Ogden Centre West gave participants the opportunity to network and engage in insightful discussions while enjoying food and refreshments, successfully fostering collaboration and connections.
The success of this event and the high level of interest—evidenced by it being oversubscribed—has paved the way for its prompt reiteration.
