| Christian Hakert.
Memory Access Analysis and Endurance Leveling Approaches for Non-volatile Working Memory Systems.
Master's Thesis, 2019
[BibTeX][PDF][Abstract]@mastersthesis { HakertMA,
title = {Memory Access Analysis and Endurance Leveling Approaches for Non-volatile Working Memory Systems},
author = {Hakert, Christian},
school = {TU Dortmund},
year = {2019},
keywords = {nvm-oma},
file = {https://ls12-joomla-host.cs.tu-dortmund.de/daes/media/documents/theses/2019-hakert.pdf},
confidential = {n},
abstract = {Emerging technologies for non-volatile memory (NVM), such as phase-change memory (PCM), ferroelectric RAM (FeRAM), spin-transfer torque magnetoresistive RAM (STT-MRAM) and many more, have been considered as a replacement for DRAM and storage due to low leakage power, high capacity and fast access times. A major disadvantage of some NVMs is the significantly lower write endurance compared to DRAM. To tackle this issue, several in-memory wear-leveling approaches have been proposed in literature, targeting the program execution to make the memory write usage more uniform. These approaches operate on different granularities, like memory pages, cache-lines, big segments, etc. Aging-aware approaches take the current wear-level of memory cells into account to make the best wear-leveling decisions, while other approaches try to balance the wear-level in a random based manner or in a circular manner. However, most approaches propose specialized hardware controllers to collect the aging information and to perform wear-leveling actions, like relocations of memory regions.
As specialized hardware may be hard to build at all, this thesis proposes so-called software only wear-leveling, which only makes use of commonly available hardware components. Aging-aware wear-leveling on the granularity of virtual memory pages is achieved by estimating the cell aging by a statistical runtime approximation. As an extension to this, fine-grained wear-leveling for the stack region of applications is achieved by relocating the stack periodically in a circular manner through a reserved memory region.
The necessity for commonly available hardware components only allows the techniques to be executed in a full system simulation setup (i.e. gem5 as a simulator for an ARM CPU and NVMain 2.0 as a simulator for the NVM) for evaluation purposes. The results show that coarse-grained aging-aware wear-leveling with approximated cell ages works out, but cannot resolve non-uniform write distributions within virtual memory pages. This limits the gained improvement of the memory lifetime up to a factor of ≈ 13 compared to the baseline without any wear-leveling. Combining the coarse-grained approach with the fine- grained stack wear-leveling technique, the non-uniformity caused by the stack region can be resolved and the evaluations result in an improvement of the memory lifetime up to a factor of ≈ 900 compared to the baseline.},
adviser = {Dr. Kuan-Hsun Chen},
}
Emerging technologies for non-volatile memory (NVM), such as phase-change memory (PCM), ferroelectric RAM (FeRAM), spin-transfer torque magnetoresistive RAM (STT-MRAM) and many more, have been considered as a replacement for DRAM and storage due to low leakage power, high capacity and fast access times. A major disadvantage of some NVMs is the significantly lower write endurance compared to DRAM. To tackle this issue, several in-memory wear-leveling approaches have been proposed in literature, targeting the program execution to make the memory write usage more uniform. These approaches operate on different granularities, like memory pages, cache-lines, big segments, etc. Aging-aware approaches take the current wear-level of memory cells into account to make the best wear-leveling decisions, while other approaches try to balance the wear-level in a random based manner or in a circular manner. However, most approaches propose specialized hardware controllers to collect the aging information and to perform wear-leveling actions, like relocations of memory regions.
As specialized hardware may be hard to build at all, this thesis proposes so-called software only wear-leveling, which only makes use of commonly available hardware components. Aging-aware wear-leveling on the granularity of virtual memory pages is achieved by estimating the cell aging by a statistical runtime approximation. As an extension to this, fine-grained wear-leveling for the stack region of applications is achieved by relocating the stack periodically in a circular manner through a reserved memory region.
The necessity for commonly available hardware components only allows the techniques to be executed in a full system simulation setup (i.e. gem5 as a simulator for an ARM CPU and NVMain 2.0 as a simulator for the NVM) for evaluation purposes. The results show that coarse-grained aging-aware wear-leveling with approximated cell ages works out, but cannot resolve non-uniform write distributions within virtual memory pages. This limits the gained improvement of the memory lifetime up to a factor of ≈ 13 compared to the baseline without any wear-leveling. Combining the coarse-grained approach with the fine- grained stack wear-leveling technique, the non-uniformity caused by the stack region can be resolved and the evaluations result in an improvement of the memory lifetime up to a factor of ≈ 900 compared to the baseline.
|
