Proktive hantering av minneshierarkier
Tidsperiod: 2015-01-01 till 2018-12-31
Projektledare: David Black-Schaffer
Budget: 3 640 000 SEK
The memory hierarchy is critical for a computer´s performance and power efficiency. However, today´s memory hierarchies are reactive and brute-force, repeatedly replicating and moving data, resulting in significant energy and performance losses. Proactively managing the memory hierarchy can dramatically reduce the energy (by moving data to the right place the first time) and improve performance (by more efficiently using the cache space). This project will investigate proactively managing the memory hierarchy by building upon our breakthrough cache designs [MICRO13, ISCA14], which reduce energy (-60%) and latency (-40%) by eliminating cache tags and providing a metadata table to locate data. We will extend the metadata to enable proactive data management policies for installation and eviction. Uniquely, the underlying design provides a unifying framework to address the full hierarchy and incorporate other designs and advances.We will use this framework to: 1) investigate proactively moving data to eliminate the wasted energy and space of current designs, 2) explore the tradeoff between associativity and number of cache levels to better match data behavior to storage, and 3) to dynamically adjust the granularity of data, thereby improving sparse data utilization. The unique framework for this project enables all three of these approaches to be undertaken together, making it possible to integrate such disparate optimizations for the first time.