Dr. Karsten Schwan (Advisor, School of Computer Science, Georgia Tech)

Dr. Umakishore Ramachandran (School of Computer Science, Georgia Tech)

Dr. Moinuddin Qureshi (School of Electrical and Computer Engineering, Georgia Tech)

Dr. Remzi H. Arpaci-Dusseau (Dept. of Computer Science, University of Wisconsin-Madison)

Dr. Greg Eisenhauer(School of Computer Science, Georgia Tech)

Abstract:
To address the 'memory wall' problem of future systems, vendors are creating heterogeneous 

memory structures, supplementing DRAM with on-chip stacked 3D-RAM and high capacity 

non-volatile memory (NVM). Each of these technologies differs significantly in terms of density, 

bandwidth, and latency. However, current operating systems (OSes) and 

software stacks lack generic memory abstractions that can be uniformly used with different 

memory types. This increases the software complexity resulting in limited performance and 

efficiency benefits from the memory heterogeneity for both virtualized and non-virtualized 

systems. To address these challenges,  this thesis develops HeteroMem -- an OS design 

for heterogeneous memory and makes the following contribution.

First, HeteroMem introduces a unified OS abstraction for heterogeneous memories. 

As a result, different memory technologies can extensively leverage the current advances 

made for traditional memory management, thereby reducing software complexity, achieving 

efficient use of hardware resources such as caches and TLBs, and permitting seamless 

scaling across heterogeneous memory components. Second, HeteroMem incorporates novel 

memory placement mechanisms focused on reducing data movement overheads. The 

outcome is up to 2x improvement in application performance compared to the 

state-of-the-art solutions.

Furthermore, to exploit the persistence benefits from non-volatile memories,  HeteroMem goes 

beyond memory capacity scaling, to provide fast persistent object storage. Using NVMs for 

persistence leads to new types of cache sharing and energy bottlenecks. We address 

these bottlenecks via novel cache- and energy-efficient system software principles that do not 

impact application correctness. Finally, for achieving maximum performance and reliability 

gains with heterogeneous memory, we also explore the redesign of HPC, datacenter, and 

end-user applications.