Hidenori Nakazato

Divisible Load Scheduling with Result Collection on Heterogeneous Systems

Divisible load theory (DLT) is an established mathematical framework to study divisible load scheduling (DLS). However, traditional DLT does not comprehensively deal with the scheduling of results back to source (i.e., result collection) on heterogeneous systems. In this paper, the DLSRCHETS (DLS with result collection on hetero- geneous systems) problem is addressed. The few papers to date that have dealt with DLSRCHETS, proposed simplistic LIFO (last in, first out) and FIFO (first in, first out) type of schedules as solutions to DLSRCHETS. In this paper, a new heuristic algorithm, ITERLP, is proposed as a solution to the DLSRCHETS problem. With the help of simulations, it is proved that the performance of ITERLP is significantly better than existing algorithms.

Model for time-varying quality of speech services

Speech services are widely deployed on new communication infrastructures. However, compared with traditional infrastructures, the ever varying conditions are native of these new infrastructures. Modeling the time-varying speech quality becomes an urgent issue. This paper proposes a computable model for time-varying speech quality evaluation, called M-Model. Considering the time-varying issue, EWMA is used in the model. M-Model is validated by a subjective experiment. The model indicates that time-varying quality with extremely high frequency does not affect perceived quality at all. In addition, we observed that the recency effect influences not only the perceived overall quality but also the perceived instantaneous quality.

Clustering-Based Task Scheduling in a Large Number of Heterogeneous Processors

Parallelization paradigms for effective execution in a Directed Acyclic Graph (DAG) application have been widely studied in the area of task scheduling. Schedule length can be varied depending on task assignment policies, scheduling policies, and heterogeneity in terms of each processor and each communication bandwidth in a heterogeneous system. One disadvantage of existing task scheduling algorithms is that the schedule length cannot be reduced for a data intensive application. In this paper, we propose a clustering-based task scheduling algorithm called Clustering for Minimizing the Worst Schedule Length (CMWSL) to minimize the schedule length in a large number of heterogeneous processors. First, the proposed method derives the lower bound of the total execution time for each processor by taking both the system and application characteristics into account. As a result, the number of processors used for actual execution is regulated to minimize the Worst Schedule Length (WSL). Then, the actual task assignment and task clustering are performed to minimize the schedule length until the total execution time in a task cluster exceeds the lower bound. Experimental results indicate that CMWSL outperforms both existing list-based and clustering-based task scheduling algorithms in terms of the schedule length and efficiency, especially in data-intensive applications.

Analysis of Divisible Load Scheduling with Result Collection on Heterogeneous Systems

Divisible Load Theory (DLT) is an established framework to study Divisible Load Scheduling (DLS). Traditional DLT ignores the result collection phase, and specifies no solution to the general case where both the network speed and computing capacity of the nodes are heterogeneous. In this paper, the DLS with Result Collection on HETerogeneous Systems (DLSRCHETS) problem is formulated as a linear program and analyzed. The papers to date that have dealt with result collection, proposed simplistic LIFO (Last In, First Out) and FIFO (First In, First Out) type of schedules as solutions. The main contributions of this paper are: (a) A proof of the Allocation Precedence Condition, which is inconsequential in LIFO or FIFO, but is important in a general schedule. (b) A proof of the Idle Time Theorem, which states that irrespective of whether load is allocated to all available processors, in the optimal solution to the DLSRCHETS problem, at the most one processor that is allocated load has idle time, and that the idle time exists only when the result collection begins immediately after the completion of load distribution.

SPORT: An algorithm for Divisible Load Scheduling with result collection on heterogeneous Systems

Divisible Load Theory (DLT) is an established mathematical framework to study Divisible Load Scheduling (DLS). However, traditional DLT does not address the scheduling of results back to source (i.e., result collection), nor does it comprehensively deal with system heterogeneity. In this paper, the DLSRCHETS (DLS with Result Collection on HET-erogeneous Systems) problem is addressed. The few papers to date that have dealt with DLSRCHETS, proposed simplistic LIFO (Last In, First Out) and FIFO (First In, First Out) type of schedules as solutions to DLSRCHETS. In this paper, a new polynomial time heuristic algorithm, SPORT (System Parameters based Optimized Result Transfer), is proposed as a solution to the DLSRCHETS problem. With the help of simulations, it is proved that the performance of SPORT is significantly better than existing algorithms. The other major contributions of this paper include, for the first time ever, (a) the derivation of the condition to identify the presence of idle time in a FIFO schedule for two processors, (b) the identification of the limiting condition for the optimality of FIFO and LIFO schedules for two processors, and (c) the introduction of the concept of equivalent processor in DLS for heterogeneous systems with result collection.

A New Token Based Protocol for Group Mutual Exclusion in Distributed Systems

In this paper we present a new token based protocol for group mutual exclusion in distributed systems. The protocol uses one single token to allow multiple processes to enter the critical section for a common session. One of the significant characteristics of the protocol is - concurrency, throughput and waiting time can be regulated adjusting the time period for which a session is declared. The minimum and the maximum number of messages to enter the CS is 0 and (n + 2) respectively where n is the total number of processes in the system. Moreover, simulation results show that the protocol, on average case, considerably reduces the number of messages per entry to the CS and also requires much lower waiting times. The maximum concurrency the protocol supports is n. The protocol also ensures no starvation in the system. Furthermore, this algorithm works out for the extended group mutual exclusion problem as well

ORC-GPS: Output Rate-Controlled Scheduling Policy for Delay Guarantees

Recently packet scheduling algorithms such as Packetized GPS (PGPS), worst-case fair weighted fair queueing (WF2Q) and Self-Clocked Fair Queueing (SCFQ) have been proposed in order to guarantee deterministic or statistical delay bounds. These algorithms are based on generalized processor sharing (GPS) which is an ideal scheduling algorithm based on fluid flow model in which the traffic is infinitely divisible. GPS provides a minimum guaranteed service rate for each session and tight delay bounds for leaky bucket constrained sessions. However, the delay bounds are unnecessarily large because each session is served according to its associated constant weight until the session buffer is empty. In this paper, we present a scheduling policy called output rate-controlled generalized processor sharing (ORC-GPS). ORC-GPS is a rate-based scheduling like GPS and controls the service rate to lower the delay bounds for leaky bucket constrained sessions. In numerical experiments, we compare ORC-GPS with GPS in terms of delay bounds.

Popularity Proportional Cache Size Allocation policy for video delivery on CCN

Content Centric Networking (CCN), which is regarded as an innovational architecture of future network, is increasingly gaining attention in recent years. Based on the observation of current Internet usage, CCN presents a content-centric model which focuses on data distribution and retrieval. To enhance the efficiency of content distribution, In-network caching is adopted, which enables routers to temporarily store the passing by data. Thus the scheme of cache management will affect the network performance significantly. In this paper, a cache replacement policy named Popularity Proportional Cache Size Allocation (PPCSA) policy has been proposed, which designed for video content delivery on CCN. Unlike other previous works in this area, which have focused on the popularity of individual content chunks, in this study, we not only consider the title-level popularity of a video which is composed of a set of video segments, but also take the chunk-level probability of video segments within a video title into consideration. We evaluated our policy through simulations, by using a simple 4-level tree topology and a real network based hybrid topology. The results validate that on both scenarios, the proposed approach can increase the average cache hit ratio and shorten the average content delivery distance between requesters and object contents. In addition, we analyzed the energy efficiency of PPCSA policy, and we found that compared to other three cache replacement policies, Least Recently Used (LRU), Least Frequently Used (LFU) and First in First out (FIFO), PPCSA reduced the transmission energy consumption, and improved the energy efficiency in idle time.

Dynamic Scheduling for Speculative Execution to Improve MapReduce Performance in Heterogeneous Environment

MapReduce framework allows users to quickly develop big-data applications and process big-data effectively. However, unexpected malfunction may be found in cloud environment because a distributed system consists of several hardware, and this malfunction often causes delay of overall processing. MapReduce framework provides Speculative Execution (SE). SE reduces delay in a homogeneous environment by assigning delayed tasks to additional nodes. As cloud computing prevails, cloud computing environment is moving from homogeneous to heterogeneous. Original SE is not perfect and sometimes produces inefficient result in a heterogeneous environment. This paper proposes Dynamic Scheduling for Speculative Execution (DSSE) which enhances performance in a heterogeneous environment by improving existing SE. DSSE prevents wasted SE since it calculates processing capability of each node more objectively and precisely. DSSE has reduced entire processing time approximately 10% compared to original SE. Success rate of SE was 100%.

Peer-to-Peer Content Distribution in Clustered Topologies with Source Coding

Network coding has been applied successfully in peer-to-peer systems to shorten the distribution time, particularly in extreme conditions where peers in the network are clustered and separated by physical links with limited bandwidth. In this paper, we focus more closely on the use of source coding, i.e. encoding is done only at the source, as an alternative to network coding to facilitate content distribution under such limited bandwidth configuration. We observe analytically and experimentally that, in this specific case, with appropriately chosen expansion factors, source coding can have comparable performance to network coding in terms of distribution time, yet consuming much less computational resources than the latter approach does.