A5035654161- Advanced Data Storage Technologies
- Caching and Content Delivery
- Peer-to-Peer Network Technologies
- Cloud Computing and Resource Management
- Parallel Computing and Optimization Techniques
- Cloud Data Security Solutions
- Distributed and Parallel Computing Systems
- Software-Defined Networks and 5G
- Nonlinear Optical Materials Studies
- Photonic Crystals and Applications
- Artificial Intelligence in Games
- Interconnection Networks and Systems
- Image and Video Quality Assessment
- Advanced Manufacturing and Logistics Optimization
- Digital Games and Media
- Computer Graphics and Visualization Techniques
- Ferroelectric and Negative Capacitance Devices
- Advanced Neural Network Applications
- Optimization and Search Problems
- Digital and Traditional Archives Management
- Photorefractive and Nonlinear Optics
- Modular Robots and Swarm Intelligence
- Research Data Management Practices
- Laser Material Processing Techniques
- Complex Network Analysis Techniques
Microsoft Research (United Kingdom)
2014-2024
Microsoft (United States)
2014
Institut national de recherche en informatique et en automatique
2009-2012
Sorbonne Université
2009-2011
Laboratoire de Recherche en Informatique de Paris 6
2011
Centre National de la Recherche Scientifique
2009-2010
Université Paris Cité
2009
AI clusters today are one of the major uses High Bandwidth Memory (HBM). However, HBM is suboptimal for workloads several reasons. Analysis shows overprovisioned on write performance, but underprovisioned density and read bandwidth, also has significant energy per bit overheads. It expensive, with lower yield than DRAM due to manufacturing complexity. We propose a new memory class: Managed-Retention (MRM), which more optimized store key data structures inference workloads. believe that MRM...
To match the blooming demand of generative AI workloads, GPU designers have so far been trying to pack more and compute memory into single complex expensive packages. However, there is growing uncertainty about scalability individual GPUs thus clusters, as state-of-the-art are already displaying packaging, yield, cooling limitations. We propose rethink design scaling clusters through efficiently-connected large Lite-GPUs, with single, small dies a fraction capabilities larger GPUs. think...
A significant fraction of data stored in cloud storage is rarely accessed. This referred to as cold data; cost-effective for has become a challenge providers. Pelican rack-scale hard-disk based unit designed the basic building block exabyte scale data. In Pelican, server, power, cooling and interconnect bandwidth resources are provisioned by design support workloads; this right-provisioning significantly reduces Pelican's total cost ownership compared traditional disk-based storage.Resource...
DHT-based P2P systems provide a fault-tolerant and scalable means to store data blocks in fully distributed way. Unfortunately, recent studies have shown that if connection/disconnection frequency is too high, may be lost. This true for most of the current systems' implementations. To deal with this problem, it necessary build more efficient replication maintenance mechanisms. In article, we study effect churn on PAST, an existing system. We then propose solutions enhance tolerance evaluate...
Massively Multiplayer Online Games (MMOGs) recently emerged as a popular class of applications with millions users. To offer acceptable gaming experience, such need to render the virtual world surrounding player very low latency. However, current state-of-the-art MMOGs based on peer-to-peer overlays fail satisfy these requirements. This happens because avatar mobility implies many data exchanges through overlay. As do not anticipate this mobility, needed is delivered time, which leads...
Sustainable and cost-effective long-term storage remains an unsolved problem. The most widely used technologies today are magnetic (hard disk drives tape). They use media that degrades over time has a limited lifetime, which leads to inefficient, wasteful, costly solutions for long-lived data. This paper presents Silica: the first cloud system archival data underpinned by quartz glass, extremely resilient allows be left in situ indefinitely. hardware software of Silica have been co-designed...
Sustainable and cost-effective long-term storage remains an unsolved problem. The most widely used technologies today are magnetic (hard disk drives tape). They use media that degrades over time has a limited lifetime, which leads to inefficient, wasteful, costly solutions for storing long-lived data. We building Silica: the first cloud system archival data underpinned by quartz glass, extremely resilient with virtually unlimited lifetime. Data is written using ultrafast laser...
Sustainable and cost-effective long-term storage remains an unsolved problem. The most widely used technologies today are magnetic (hard disk drives tape). They use media that degrades over time has a limited lifetime, which leads to inefficient, wasteful, costly solutions for long-lived data. This paper presents Silica: the first cloud system archival data underpinned by quartz glass, extremely resilient allows be left in situ indefinitely. hardware software of Silica have been co-designed...
The demand for long-term data storage in the cloud grows continuously into zettabytes. Operating at such scales requires a fundamental re-thinking of how we build large-scale systems to archive sustainable and costeffective manner. In Project Silica, technology is being designed developed from media up by leveraging recent progress ultrafast laser nano-structuring transparent media. Together with advances reading, decoding error correction processes, high-density high-throughput...
Large-scale distributed systems gather thousands of peers spread all over the world. Such need to offer good routing performances regardless their size and despite high churn rates. To achieve that requirement, system must add appropriate shortcuts its logical graph (overlay). However, choose efficient shortcuts, obtain information about overlay topology. In case heterogeneous peer distributions, retrieving such is not straightforward. Moreover, due churn, topology rapidly evolves, making...
Recent progress in ultrafast laser nanostructuring enables high-density multidimensional volumetric data writing. With exceptional media longevity, this could transform archival cloud storage. We demonstrate peak densities above 250 MB/mm 3 of write and read.