Talks

Sistemas Heterogeneous de Memoria

Conference: Yachay Tech

Year: 09/08/2023

Description:

In this talk we touched on the evolution of computing systems and how they have been affected by the heterogeneity of the systems. This specifically has changed memory systems a lot, so it has been necessary to take certain precautions and considerations for the development and modification of HPC applications.

Link: 

Performance Metrics for Managing Heterogeneous Memory in HPC Applications

Conference: ADAC Seminar

Year: 28/04/2023

Description:

Over the past decade, memory systems have become increasingly complex, with supercomputers now featuring multiple levels of heterogeneous and non-uniform memory, each with unique performance, capacity, bandwidth, latency, and power consumption characteristics. This poses a significant challenge for developers of scientific applications, who must efficiently utilize the memory system to optimize performance. Firstly, we propose a method to simplify the management of modern memory systems by categorizing memory types and describing their hardware characteristics using metrics such as bandwidth, latency, and capacity. This allows runtime systems, parallel libraries, and scientific applications to select the most appropriate memory for each allocation without needing to modify the code for each platform. Secondly, power performance metrics was analyzed by evaluating scenarios where developers need to select the most power-efficient memory for their executions or balance application performance and power consumption.

Link: 

Analysing power performance on Heterogeneous Memory Systems

Conference: R-CCS Café

Year: 2023

Description:

Memory systems in High Performance Computing systems have undergone major changes in recent years. The trend was to have main memory, storage, and multiple levels of cache, but now servers come with non-uniform memory access and can contain multiple types of memory.

This work concentrates efforts on the issue of the use of various types of memory with different properties; we've specifically taken into account the power consumption metrics of each memory and decided where to allocate buffers when developers face power-constraining situations. In the same way, we also provide an early simulation model for those developers who don't have access to HMS and want to get an idea of how much power their applications could use on a given system.

This talk provides an overview of HMS, summarizes the various results, and provides a future perspective.

Link: 

Identifying suitable Storage Target for Surrogate Model Training on Heterogeneous Storage Infrastructures

Conference: CEA-RIKEN Workshop 2023-1

Year: 2023

Description:

This talk updates the activities on accelerate training on surrogates by better utilise the storage system throught a methodology that exposes a relation between storage performance and AI metrics which can have some affinities to certain storage kinds.

Link: 

Accelerating convergence of simulation and AI with hierarchical memory/storage systems

Conference: CEA-RIKEN Workshop 2022-2

Year: 2022

Description:

This talk proposes the acceleration of the interaction between model simulation and surrogate models (training) throught better utilise the memory and storage system.

Link: 

Using Performance Attributes for Managing Heterogenelous Memory in HPC

Conference: PDSEC - IPDPS22

Year: 2022

Description:

The complexity of memory systems has increased considerably over the past decade. Supercomputers may now include several levels of heterogeneous and non-uniform memory, with significantly different properties in terms of performance, capacity, persistence, etc. Developers of scientific applications face a huge challenge: efficiently exploit the memory system to improve performance, but keep productivity high by using portable solutions. In this work, we present a new API and a method to manage the complexity of modern memory systems...

Link: https://hal.inria.fr/hal-03599360/document

Determining Criteria for Data Allocation in Heterogeneous Memory Systems for HPC Applications

Conference: The 4th R-CCS International Symposium

Year: 2022

Description:

The complexity of memory systems has increased in the last few years. Supercomputers may include heterogeneous memory systems (HMS), with very different kinds of memory in terms of performance, capacity, latency, bandwidth, etc. In that sense, developers of applications have a new challenge when trying to better exploit the memory system...