Title: OS Support for Heterogeneous Memory Management

Sudarsun Kannan

School of Computer Science

College of Computing

Georgia Institute of Technology

Date: Friday, December 11th, 2015

Time: 3 PM to 5 PM EST

Location: KACB 3402

Committee:

------------

Dr. Karsten Schwan (Advisor, School of Computer Science, Georgia Tech) Dr. Ada Gavrilovska (Advisor, School of Computer Science, Georgia Tech) Dr. Umakishore Ramachandran (School of Computer Science, Georgia Tech) Dr. Sudhakar Yalamanchili (School of Electrical and Computer Engineering, 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)

Abstract:

-----------

To address the 'memory wall' problem of future datacenter servers, 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. Consequently, future systems are expected to address these requirements with heterogeneous memory architectures. However, memory heterogeneity will increase the software/OS management complexity. Therefore, a unified OS-level abstraction and efficient management principles are required for maximizing application benefits in both virtualized and non-virtualized systems. To date, operating systems, or hypervisors, do not provide a unified memory abstraction, and are ineffective in dealing with memory heterogeneity. To address these challenges, this thesis proposes an OS design 'HeteroMem' for heterogeneous memory support and makes the following contribution.

First, the thesis proposes a unified NUMA-like abstraction for different heterogeneous memory resources for seamless scaling and scalable OS-level management. The abstraction is exposed to applications as well as other OS subsystems. Second, HeteroMem builds customized managers specific to memory characteristics such as capacity, bandwidth, and latency, by extending the OS virtual memory

(VM) subsystem. Extending the VM subsystem enables seamless scaling across different memory types, and also inherits the virtues of VM such as efficient cache and TLB use. Third, HeteroMem exploits the OS-level information about applications for intelligent data placement with a focus on reducing data movement overheads. Fourth, HeteroMem supports memory heterogeneity in virtualized environment by exposing different memory types to virtual machines (guest-OSs), and designing a scalable coordinated hypervisor-guest OS memory management.

Further, to exploit the persistence benefits in memory types such as NVM, this thesis extends beyond memory capacity scaling to provide fast persistent storage as well as support fault tolerance mechanisms such as checkpointing. Supporting persistence is expensive in terms of processor cache sharing impact, and energy overheads. This thesis also investigates efficient system software principles to reduce such overheads.