Cross-layer resilience for rack-scale disaggregated memory

Descripción del proyecto

Más memoria para las aplicaciones con un uso intensivo de datos

El rápido aumento de la conectividad social y del comercio electrónico requiere una memoria informática cada vez mayor para almacenar y procesar mayores cantidades de datos con una baja latencia. Sin embargo, que un ordenador tenga más memoria significa que habrá más errores de memoria, lo que puede interrumpir la disponibilidad del servicio y causar pérdidas de datos. En este contexto, el proyecto financiado con fondos europeos SHEPHERD desarrollará técnicas de resiliencia de la memoria entre capas en las que el «software» y el «hardware» trabajan juntos para evitar la indisponibilidad del servicio y las pérdidas de datos debido a errores de memoria en la memoria desagregada con una baja sobrecarga. El proyecto, en forma de beca de investigación, aprovechará los puntos fuertes del investigador experimentado en el ámbito del diseño de «software» de sistemas para la memoria no volátil y los conocimientos especializados del anfitrión en el ámbito de la tolerancia a los fallos de memoria y las arquitecturas de microservidores.

Objetivo

Data-intensive applications, such as search, social connectivity, and e-commerce, are increasingly relying on computer memory for storing and processing a large amount of data with low latency. Emerging disaggregated memory meets this growing demand by leveraging recent technology advances in high-density, byte-addressable non-volatile memory and low-latency interconnects to provide a large memory pool shared across micro-server compute nodes. With a larger memory density and capacity though, memory errors will become more frequent. Addressing such errors becomes paramount as they can cripple service availability and cause data loss. However, established hardware techniques for tolerating memory errors transparently to application software, originally developed for DRAM-based main-memory systems, become impractical at the scale of disaggregated memory due to increasing performance, energy and area overheads. This research will investigate and develop cross-layer memory resilience techniques where software and hardware work together cooperatively to prevent service unavailability and data loss due to memory errors in disaggregated memory with low overhead. It capitalizes on the existing strengths of the experienced researcher (ER) in the area of system software design for non-volatile memory and expertise of the host on memory fault tolerance and micro-server architectures. It will lead to several publications in top-quality venues and will inform forthcoming industrial efforts to develop resilient disaggregated memory micro-server architectures, thus contributing to the research output, visibility, and economy of Europe. Overall, the fellowship will give the ER the opportunity to reintegrate into a long-term academic position in Europe, after a long and fruitful research career in the academic and industrial sectors of the United States. The ER will emerge from the project with new skills and the capability to effectively lead his own research group.