• KIPS Transactions on Computer and Communication Systems
• /
• v.3 no.11
• /
• pp.409-418
• /
• 2014

The amount of data located in storage servers has dramatically increased with the growth in cloud and social networking services. Storage systems with very large capacities may suffer from poor reliability and long latency, problems which can be addressed by the use of a hybrid disk, in which mechanical and flash memory storage are combined. The Linux-based SSD(solid-state disk) uses a caching technique based on the DM-cache utility. We assess the limitations of DM-cache by evaluating its performance in diverse environments, and identify problems with the caching policy that it operates in response to various commands. This policy is effective in reducing latency when Linux is running in native mode; but when Linux is installed as a guest operating systems on a virtual machine, the overhead incurred by caching actually reduces performance.

• Journal of Internet Computing and Services
• /
• v.11 no.4
• /
• pp.11-22
• /
• 2010

In case that group communication protocols assume every process is interested in all events occurring in a large scale group, the events multicasting to a subset of a large process group, such as a sensor network, potentially varying for every event based on their interests might lead to very high communication overhead on each individual process. Moreover, despite the importance of both guaranteeing message delivery order and supporting overlapped group using gossip based group communication for multicasting in sensor or P2P networks, there exist little research works on development of gossip-based protocols to satisfy all these requirements. In this paper, we present a new gossip-based causal message order guaranteeing multicast protocol based on local views and delegates representing subgroups and fully utilizing multi-group features to improve scalability. In the proposed protocol, the message delivery order in overlapped groups has been guaranteed consistently by all corresponding members of the groups including delegates. Therefore, these features of the proposed protocol might be significantly scalable rather than those of the protocols guaranteeing atomic order dependencies between multicast messages by hierarchical membership list of dedicated groups like traditional committee protocols and much stronger rather than fully decentralized protocols guaranteeing dependencies between multicast messages based on only local views. And the proposed protocol is a hybrid approach improving the inherent scalability of multicast nature by gossip-based technique in all communications.

• Journal of the Institute of Electronics Engineers of Korea SD
• /
• v.40 no.5
• /
• pp.345-353
• /
• 2003

When designing network processors or implementing network equipments such as routers are implemented, IP lookup operations cause the major impact on their performance. As the organization of the IP address becomes simpler, the speed of the IP lookup operations can go faster. However, since the efficient management of IP address is inevitable due to the increasing number of network users, the address organization should become more complex. Therefore, for both IPv4(IP version 4) and IPv6(IP version 6), it is the essential fact that IP lookup operations are difficult and tedious. Lots of researcher for improving the performance of IP lookups have been presented, but the good solution has not been came out. Software approach alleviates the memory usage, but at the same time it si slow in terms of searching speed when performing an IP lookup. Hardware approach, on the other hand, is fast, however, it has disadvantages of producing hardware overheads and high memory usage. In this paper, conventional researches on IP lookups are shown and their advantages and disadvantages are explained. In addition, by mixing two representative structures, a new hybrid parallel architecture for fast IP lookups is proposed. The performance evaluation result shows that the proposed architecture provides better performance and lesser memory usage.

• KIPS Transactions on Computer and Communication Systems
• /
• v.10 no.7
• /
• pp.191-198
• /
• 2021

Since the size of training dataset become large and the model is getting deeper to achieve high accuracy in deep learning, the deep neural network training requires a lot of computation and it takes too much time with a single node. Therefore, distributed deep learning is proposed to reduce the training time by distributing computation across multiple nodes. In this study, we propose hybrid allreduce strategy that considers the characteristics of each layer and communication and computational overlapping technique for synchronization of distributed deep learning. Since the convolution layer has fewer parameters than the fully-connected layer as well as it is located at the upper, only short overlapping time is allowed. Thus, butterfly allreduce is used to synchronize the convolution layer. On the other hand, fully-connecter layer is synchronized using ring all-reduce. The empirical experiment results on PyTorch with our proposed scheme shows that the proposed method reduced the training time by up to 33% compared to the baseline PyTorch.