Evaluating the programmability and scalability of memory hierarchies with read-only data blocks
Master of Science
This thesis evaluates the programmability and scalability of Fresh Breeze Memory (FBM), a unique memory hierarchy design with universally identifiable fixed-size read-only blocks. As a part of this work, we implemented an FBM emulator capable of emulating each level of the FBM hierarchy. Using the results obtained from this emulator, we show that systems with such a memory hierarchy can be programmed using task structured parallel programming languages such as Habanero Java (HJ), and can yield scalable cache performance. Our comparative study of the FBM and conventional memory architectures indicates that the FBM’s performance can scale out to many cores by avoiding the coherence bottleneck of conventional memory systems. Lastly, our results indicate that some data layouts that are known to perform well in conventional memory hierarchies also interact well with the FBM.