| 1. Cluster |
| 1.1 Coscheduling |
| Batch scheduling is the most popular
policy to manage dedicated clusters for running non-interactive jobs.
The gang scheduling (Explicit Coscheduling) is an efficient
co-scheduling algorithm for fine-grained parallel processes. Unlike the
gang scheduling, with a communication-driven coscheduling like Dynamic
Coscheduling, Spin-Block, Periodic Boost or Co-ordinated Coscheduling,
each node in a cluster has an independent scheduler, which coordinates
the communicating processes of a parallel job. All these coscheduling
algorithms rely primarily on one of the two local events (arrival of
a message and waiting for a message) to determine when and
which process to schedule. |
| 1.2 Data Center |
|
Design of high performance,
scalable and dependable data centers has become a critical issue to cope
up with the increasing use of the Internet in supporting various
Web-based services. As the network bandwidth is improving faster than
the server capacity, data centers are anticipated to be the bottleneck
in hosting network-based services. It has been observed that Web
servers/data centers contribute to about 40% of the overall delay and
this delay is likely to grow with the increasing use of dynamic Web
contents. Although several hardware and software scale-up techniques
have been proposed to enhance the performance of single node servers,
these are not viable/long-term solutions considering the fact that
online users are growing at about 90% annually. In view of this,
cluster-based data centers are envisioned to answer most of the
requirements in a cost-effective manner. |
| |
| 2.
Interconnection Networks |
| 2. 1 System Area
Network |
| For designing scalable, high performance
computing in a variety of distributed and parallel systems, this group
explores QoS capable SAN models with admission & congestion control,
fault tolerant system, and energy optimization techniques. The
comprehensive simulation testbed has been developed for these issues. |
| 2.2 Network on Chip |
|
System-on-Chip (SoC) architecture are being dramatically
applied in an emerging and exciting research field. Therefore, design of
scalable, adaptable, high performance, area and energy efficient on-chip
networks are crucial to the success of SoC design and embedded systems. |
|
|
|
3.
Wireless Network |
|
The overall research is to design and
analyze wireless and mobile networks that can provide QoS guarantees and
high communication performance. In particular, a unified approach for
QoS provisioning in cellular networks is investigated to provide
improved and predictable performance. Second, the design of Internet
based mobile ad hoc networks (IMANETs) is investigated for providing
universal data accessibility. Next, since updating a cached or
replicated data item is a critical issue for supporting caching in
mobile wireless networks, an efficient cache invalidation strategy is
currently investigated. Finally, the issues related to TCP traffic over
IMANETs will be investigated. |
|
|
|
4.
QoS (Quality of Service) Provisioning in
Internet |
|
4.1
Active Queue Management Scheme |
|
Congestion can
bring significant performance degradation and compromise the QoS to both
the users and the network. One of the most active research areas to
deal with this problem is Active Queue Management
(AQM). As the network is quite dynamic by
nature, an AQM scheme without adaptability can hardly
meet the QoS requirement under changing network and
traffic conditions. In addition, most of AQM schemes are not fair in
distributing the network resources among competing users. An extreme case of
unfairness has the same effect as a denial-of-service
(DoS) attack to the population of users who deploy the standard congestion
control mechanism. This problem becomes
increasingly critical as the widely deployed open source system software
enables a malicious user to jam the network without any congestion
control mechanism being employed. Aside from
fairness, it is also important to differentiate the
QoS requirements for different users or aggregation
(class) of users. However, not all AQM schemes can be directly adapted to
support class-based services such as the
Differentiate Service (Diffserv) architecture. |
|
4.2
Extending to Wireless Network |
|
With the rapid development in
wireless technologies, it is natural to extend the Internet
infrastructure to incorporate wireless networks.
However, wireless networks differ from wired networks
in many aspects including
resource management, congestion control, and security
concerns. The existing mechanisms cannot be easily
applied to the wireless domain without significant
revision. We investigate a variety of ways to extending
current congestion control and security issues to wireless network. |
|
|
