• Journal of KIISE:Information Networking
• /
• v.30 no.3
• /
• pp.407-415
• /
• 2003

As services requesting data transfer from a sender to a multiple number of receivers become popular, efficient group communication mechanisms like multicast get much attention. Since multicast is more efficient than unicast in terms of bandwidth usage and group management for group communication, many multicast protocols providing scalability and reliability have been proposed. Recently, router-supported reliable multicast protocols have been proposed because routers have the knowledge of the physical multicast tree structure and, in this scheme, repliers which retransmit lost packets are selected by routers. Repliers are selected dynamically based on the network situation, therefore, any receiver within a multicast group can become a replier Hence, all receivers within a group maintains a buffer for loss recovery within which received packets are stored. It is an overhead for all group receivers to store unnecessary packets. Therefore, in this paper, we propose a new scheme which reduces resource usage by discarding packets unnecessary for loss recovery from the receiver buffer. Our scheme performs the replier selection and the loss recovery of lost packets based on the LSM [1] model, and discards unnecessary packets determined by ACKs from erasers which represent local groups.

• Journal of KIISE:Information Networking
• /
• v.30 no.5
• /
• pp.656-667
• /
• 2003

Congestion control is a key task in reliable multicast along with error control. However, existing tree-based congestion control schemes such as MTCP and TRAMCC are designed for one-to-many reliable multicast and have some drawbacks when they are used for many-to-many reliable multicast. We propose an efficient congestion control mechanism, TMRCC, for tree-based many-to-many reliable multicast protocols. The proposed scheme is based on the congestion windowing mechanism and a rate controller is used in addition. The feedback for error recovery is exploited for congestion control as well to minimize the overhead at the receivers. The ACK timer and the NACK timers are set dynamically reflecting the network condition changes. The rate regulation algorithm in the proposed scheme is designed to help the flows sharing the same link to achieve the fair share quickly The performance of the proposed scheme is evaluated using ns-2. The simulation results show that the proposed scheme outperforms TRAMCC in terms of intra- session fairness and shows good level of responsiveness, TCP-friendliness, and scalability. In addition, we implemented the proposed scheme by integrating with GAM that is one of many-to-many reliable multicast protocols and evaluated the performance in a laboratory-wide testbed.

• Journal of KIISE:Computer Systems and Theory
• /
• v.31 no.5_6
• /
• pp.371-378
• /
• 2004

This paper presents the implementation and performance of the HVIA-GE card, which is a hardware implementation of the Virtual Interface Architecture (VIA) based on Gigabit Ethernet. The HVIA-GE card is a 32-bit/33MHz PCI adapter containing an FPGA for the VIA protocol engine and a Gigabit Ethernet chip set to construct a high performance physical network. HVIA-GE performs virtual-to-physical address translation, Doorbell, and send/receive completion operations in hardware without kernel intervention. In particular, the Address Translation Table (ATT) is stored on the local memory of the HVIA-GE card, and the VIA protocol engine efficiently controls the address translation process by directly accessing the ATT. As a result, the communication overhead during send/receive transactions is greatly reduced. Our experimental results show the maximum bandwidth of 93.7MB/s and the minimum latency of 11.9${\mu}\textrm{s}$. In terms of minimum latency HVIA-GE performs 4.8 times and 9.9 times faster than M-VIA and TCP/IP, respectively, over Gigabit Ethernet. In addition, the maximum bandwidth of HVIA-GE is 50.4% and 65% higher than M-VIA and TCP/IP respectively.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.16 no.5
• /
• pp.531-536
• /
• 2006

Recently, many of researchers have been studied in data processing oriented middleware for wireless sensor networks with the rapid advances on sensor and wireless communication technologies. In a wireless sensor network, a middleware should handle the data loss problem at an intermediate sensor node caused by instantaneous data burstness to support efficient processing and fast delivering of the sensing data. To handle this problem, a simple data discarding or data compressing policy for reducing the total amount of data to be transferred is typically used. But, data discarding policy decreases the correctness of a collected data, in other hand, data compressing policy requires additional processing overhead with the high complexity of the given algorithm. In this paper, it proposes a data-average method for enhancing the efficiency of data aggregation and correctness where the sensed data should be delivered only with the limited computing power and energy resource. With the proposed method, unnecessary data transfer of the overlapped data is eliminated and data correctness is enhanced by using the proposed averaging scheme when an instantaneous data burstness is occurred. Finally, with the TOSSTM simulation results on TinyBB, we show that the correctness of the transferred data is enhanced.

• 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.

• The KIPS Transactions:PartC
• /
• v.10C no.5
• /
• pp.625-634
• /
• 2003

More than a decade after its initial proposal, deployment of IP Multicast has been limited due to the problem of traffic control in multicast routing, multicast address allocation in global internet, reliable multicast transport techniques etc. Lately, according to increase of multicast application service such as internet broadcast, real time security information service etc., overlay multicast is developed as a new internet multicast technology. In this paper, we describe an overlay multicast protocol and propose fast join mechanism that considers switching of the tree. To find a potential parent, an existing search algorithm descends the tree from the root by one level at a time, and it causes long joining latency. Also, it is try to select the nearest node as a potential parent. However, it can't select the nearest node by the degree limit of the node. As a result, the generated tree has low efficiency. To reduce long joining latency and improve the efficiency of the tree, we propose searching two levels of the tree at a time. This method forwards joining request message to own children node. So, at ordinary times, there is no overhead to keep the tree. But the joining request came, the increasing number of searching messages will reduce a long joining latency. Also searching more nodes will be helpful to construct more efficient trees. In order to evaluate the performance of our fast join mechanism, we measure the metrics such as the search latency and the number of searched node and the number of switching by the number of members and degree limit. The simulation results show that the performance of our mechanism is superior to that of the existing mechanism.

• Journal of KIISE:Computing Practices and Letters
• /
• v.13 no.5
• /
• pp.293-299
• /
• 2007

Intelligent pipeline inspection gauges (PIGs) are inspection vehicles that move along within a (gas or oil) pipeline and acquire signals (also called sensor data) from their surrounding rings of sensors. By analyzing the signals captured in intelligent PIGs, we can detect pipeline defects, such as holes and curvatures and other potential causes of gas explosions. There are two major data access patterns apparent when an analyzer accesses the pipeline signal data. The first is a sequential pattern where an analyst reads the sensor data one time only in a sequential fashion. The second is the repetitive pattern where an analyzer repeatedly reads the signal data within a fixed range; this is the dominant pattern in analyzing the signal data. The existing PIG software reads signal data directly from the server at every user#s request, requiring network transfer and disk access cost. It works well only for the sequential pattern, but not for the more dominant repetitive pattern. This problem becomes very serious in a client/server environment where several analysts analyze the signal data concurrently. To tackle this problem, we devise a fast in-memory cache manager, called T-Cache, by considering pipeline sensor data as multiple time-series data and by efficiently caching the time-series data at T-Cache. To the best of the authors# knowledge, this is the first research on caching pipeline signals on the client-side. We propose a new concept of the signal cache line as a caching unit, which is a set of time-series signal data for a fixed distance. We also provide the various data structures including smart cursors and algorithms used in T-Cache. Experimental results show that T-Cache performs much better for the repetitive pattern in terms of disk I/Os and the elapsed time. Even with the sequential pattern, T-Cache shows almost the same performance as a system that does not use any caching, indicating the caching overhead in T-Cache is negligible.