Page 23 - 2017 Brochure
P. 23
ious machine learning methods, graph theory, statistics, meteorological models, or Research Description
a combination thereof. However, most studies fail to make accurate predictions for an
urban areas, because the area itself produces pollution as well as receiving pollution Meng-Chang Chen
from remote sources. In this research, we consider the characteristics of urban areas
and design a deep learning architecture composed of various neural networks, including Research Fellow
autoencoders, layers of experts, LSTMs (Long Short Term Memory), RBNs (Restricted
Boltzmann machine). Currently, we are working on incorporating meteorological models Yuan-Hao Chang
to predict transport of remote pollutants in order to further improve predictions..
Associate Research Fellow
4. Indexing, Data Mining and Management for Non-Volatile Main Memory
Hong-Yuan Mark Liao
Non-volatile memory (NVM) has been instrumental in the evolution of next-generation
memory architecture and has become a popular alternative to replace DRAM as the main Distinguished Research Fellow
memory of domain-specific computing systems. In addition to its non-volatility, NVM
has very low idle power (or leakage power) and higher cell density for capacity scaling. De-Nian Yang
However, compared to DRAM, NVM usually has longer write latency and write energy.
Our research focus is to propose new designs that improve the access performance Research Fellow
of NVM-based main memory architecture and accommodate special access patterns
found in many popular internet of things (IoT) and in-memory database applications. Mi-Yen Yen
For example, we propose a new index design, called access-pattern-aware cache-
line-based tree (xB+- tree), to improve the access performance of NVM-based main Associate Research Fellow
memory architecture by considering the cache-line-based access behavior between CPU/
processor and main memory. This design will keep frequently accessed key-value pairs or
indexes in the minimal number of cache lines and aggregate consecutively queried data
items in the same cache line, minimizing the number of accesses to NVM main memory.
We implemented and evaluated the effectiveness of the design on the Gem5 simulator.
In the future, we will continue
to study index designs with
NVM main memory for big
data applications and in-
memory computing systems.
For example, in big data
applications, data may be
frequently generated and
rarely deleted, with only a
small portion of data items
being frequently queried.
Thus, we must devise
methods to increase the
cache hit ratio and decrease
the cache miss ratio by pro-
actively and smartly placing
data items into different cache
lines, according to the special access patterns of each big data application. In addition to
performance, the endurance of NVM main memory will also be investigated.
New NVMs have large capacity, low cost and are energy efficient. These technologies are
good candidates to replace traditional DRAM and enable in-memory computing. We aim
to revisit data mining algorithms, such as frequent pattern mining, and re-design these
algorithms to be NVM friendly by using more read operations than write operations. In this
way, we can leverage the advantages of NVM for in-memory data mining while minimizing
the memory access latency and enhancing its endurance.
21
a combination thereof. However, most studies fail to make accurate predictions for an
urban areas, because the area itself produces pollution as well as receiving pollution Meng-Chang Chen
from remote sources. In this research, we consider the characteristics of urban areas
and design a deep learning architecture composed of various neural networks, including Research Fellow
autoencoders, layers of experts, LSTMs (Long Short Term Memory), RBNs (Restricted
Boltzmann machine). Currently, we are working on incorporating meteorological models Yuan-Hao Chang
to predict transport of remote pollutants in order to further improve predictions..
Associate Research Fellow
4. Indexing, Data Mining and Management for Non-Volatile Main Memory
Hong-Yuan Mark Liao
Non-volatile memory (NVM) has been instrumental in the evolution of next-generation
memory architecture and has become a popular alternative to replace DRAM as the main Distinguished Research Fellow
memory of domain-specific computing systems. In addition to its non-volatility, NVM
has very low idle power (or leakage power) and higher cell density for capacity scaling. De-Nian Yang
However, compared to DRAM, NVM usually has longer write latency and write energy.
Our research focus is to propose new designs that improve the access performance Research Fellow
of NVM-based main memory architecture and accommodate special access patterns
found in many popular internet of things (IoT) and in-memory database applications. Mi-Yen Yen
For example, we propose a new index design, called access-pattern-aware cache-
line-based tree (xB+- tree), to improve the access performance of NVM-based main Associate Research Fellow
memory architecture by considering the cache-line-based access behavior between CPU/
processor and main memory. This design will keep frequently accessed key-value pairs or
indexes in the minimal number of cache lines and aggregate consecutively queried data
items in the same cache line, minimizing the number of accesses to NVM main memory.
We implemented and evaluated the effectiveness of the design on the Gem5 simulator.
In the future, we will continue
to study index designs with
NVM main memory for big
data applications and in-
memory computing systems.
For example, in big data
applications, data may be
frequently generated and
rarely deleted, with only a
small portion of data items
being frequently queried.
Thus, we must devise
methods to increase the
cache hit ratio and decrease
the cache miss ratio by pro-
actively and smartly placing
data items into different cache
lines, according to the special access patterns of each big data application. In addition to
performance, the endurance of NVM main memory will also be investigated.
New NVMs have large capacity, low cost and are energy efficient. These technologies are
good candidates to replace traditional DRAM and enable in-memory computing. We aim
to revisit data mining algorithms, such as frequent pattern mining, and re-design these
algorithms to be NVM friendly by using more read operations than write operations. In this
way, we can leverage the advantages of NVM for in-memory data mining while minimizing
the memory access latency and enhancing its endurance.
21
