Sayaka Akioka

HPC Benchmarks on Amazon EC2

Cloud computing is grabbing people’s attention rapidly as a convenient resource of computational power, and several commercial cloud computing services are accelerating the situation. While priced cloud computing services save pains to maintain the computational environment, there are several drawbacks such as overhead of virtual machines, possibility to share one physical machine with several virtual machines, and indeterminacy of topological allocation of their own virtual machines. This paper verifies usability of Amazon Elastic Computing Cloud (Amazon EC2) from the view of both value as a research tool, and cost performance as an alternative high performance computing environment to supercomputers. We evaluated computational performance through some experiments with several high performance computing benchmarks, and estimated the operational cost.

Extended forecast of CPU and network load on computational Grid

To achieve effective load balancing and a robust Grid environment, extended load forecast for computational resources is increasingly required. Thus, this paper proposes a method of predicting network and CPU load variance within a wide range, from several minutes to over a week. This is the widest range of prediction of the existing algorithms in the load of computational resources for the Grid environment. The distinctiveness of our algorithm is in using seasonal load variation for both load variance and one-step-ahead prediction. We apply seasonal fluctuation in CPU load to network load variation especially for network load variance prediction. Furthermore, the Markov model-based meta-predictor is used for one-step-ahead prediction, which is sensitive to late trends. The results of the experiments demonstrate that our algorithm gives a good curve for expected 8-day-long load variance, and makes accurate one-step-ahead predictions. The mean error rate for one-step-ahead predictions is 9.4% in the case of network load, and 6.2% in the case of CPU load. Moreover, the least mean error rate for wider range forecasts is 5.5% for network load variation, and 3.6% for CPU load variation.

Ring data location prediction scheme for Non-Uniform Cache Architectures

Increases in cache capacity are accompanied by growing wire delays due to technology scaling. Non-uniform cache architecture (NUCA) is one of proposed solutions to reducing the average access latency in such cache designs. While most of the prior NUCA work focuses on data placement, data replacement, and migration related issues, this paper studies the problem of data search (access) in NUCA. In our architecture we arrange sets of banks with equal access latency into rings. Our last access based (LAB) prediction scheme predicts the ring that is expected to contain the required data and checks the banks in that ring first for the data block sought. We compare our scheme to two alternate approaches: searching all rings in parallel, and searching rings sequentially. We show that our LAB ring prediction scheme reduces L2 energy significantly over the sequential and parallel schemes, while maintaining similar performance. Our LAB scheme reduces energy consumption by 15.9% relative to the sequential lookup scheme, and 53.8% relative to the parallel lookup scheme.

Link Shutdown Opportunities During Collective Communications in 3-D Torus Nets

As modern computing clusters used in scientific computing applications scale to ever-larger sizes and capabilities, their operational energy costs have become prohibitive. While it is an emerging trend in modern cluster design to optimize for low energy consumption in the individual computational nodes, little attention has been paid to reducing the energy used by the communication network that connects the nodes. In this work, we consider a 3D torus network similar to the one in BlueGene/L to explore opportunities for link shutdown during collective communication operations. For example, we demonstrate that in the case of all-to-one reduce codes, approximately 99% of the total network link time can be spent in a shutoff state on a 64-node toroidal network, thus reducing the overall system energy by approximately 15-28%.

Cross-media impact on twitter in japan

Twitter, a microblogging service, is now grabbing attention of people as a new channel. For deep understanding of this new service, this paper reports the characteristics of Twitter users in Japan, and the impact of media such as publications, and TV programs on Twitter community. To the best of our knowledge, this paper is the first to analyze mutual impact between Twitter, and other media quantitatively. In order for the analyses, we crawled user profiles whose language setting is Japanese, and conducted several analysis with well-known methodologies as conventional work did. We confirmed the characteristics of the collected user profiles. We observed the distributions of the number of friends, and the number of follows both follow power-law, and there exists the correlation between the number of friends, and the number of follows. Besides the collected user profiles, we also utilized closed caption data of TV programs in Japan, and other information on media picked up Twitter. We run a batch of matching these data outside Twitter with the collected user profiles, and concluded Twitter has been already widely spread among Japanese people, however, media have still huge impact on the growth of Twitter users. We also conjectured the impact is not one-sided, however, is mutual influence between Twitter, and other media.

The Markov Model Based Algorithm to Predict Networking Load on the Computational Grid

The computational Grid is currently gaining in popularity, and it enables computers scattered all over the world to be connected by the Internet as if they are part of a large computational infrastructure. While the computational Grid gathers more and more computational resources and the number of the applications for the computational Grid is increasing, load balancing for the computational Grid is still not effective enough. Because the computers are connected by a wide area network on the computational Grid, the significant communication latency and the frequency of large wave throughputs make it difficult to achieve effective load balancing. Thus, in this paper, we propose an algorithm to predict networking loads on the computational Grid to make the use of computational resources more efficient. The proposed algorithm based on the Markov model is evaluated using an actual networking load. As a result, the Markov model based algorithm offers the most accurate predictions compared with the related work.

Ring Prediction for Non-Uniform Cache Architectures

Increasing wire delays and memory capacities motivate new ways of designing 12 and 13 caches. NUCA (non-uniform cache architecture) has received considerable attention in the last few years. While most of the prior NUCA-based efforts have focused on data placement, data replacement, and migration related issues, this paper studies the problem of data search. Specifically, it proposes and experimentally evaluates several data search schemes for NUCA L2 caches that exhibit different performance-power trade-offs. These schemes are based on predicting the next ring (set of banks) to be accessed in a NUCA L2, and checking the banks in that ring first. In this work, we present the details of these prediction schemes, and compare them to two alternate approaches: searching all rings in parallel, and searching rings sequentially, starting with the one that is closest to the CPU.

Promoting consumer products with fictional stories

Our everyday consumer lifestyle has been enhanced by embedding stories in our daily life. The stories define the meaning of an artifact appeared in them. In Japan, promoting consumer products with fictional animation stories is recently very common. We believe that analyzing the stories gives us useful insights to design future ambient intelligent services that integrate virtual and real worlds. This paper discusses the analysis of several product promotions that use fictional Japanese animation movies, and presents guidelines for successful promotions. The insights presented in the paper are effective for designing future product promotions.

Information Filter for Ambiguous Information Retrieval

The quick spread of web services has triggered the flood of information, and requests people to choose a valid set of queries for useful information retrieval in order to extract what they really need. In actual, however, appropriate query generation is recognized as one of the advanced tasks, and an information retrieval system that allows ambiguous search is desired. This paper addresses this problem through the information filter with kansei engineering approach. Kansei engineering is an activity to embrace psychological needs and feelings of users into engineering field, and often requires psychological profiling. We propose a methodology to apply kansei engineering to information filter in computer scientific fashion, so as to utilize the information filter as a recommender that accepts ambiguous information search.

A Library-based Performance Tool for Multicore Pervasive Servers

This paper proposes SPLiT (Scalable Performance Library Tool) as the methodology to improve performance of applications on multicore processors through CPU and cache optimizations on the fly. SPLiT is designed to relieve the difficulty of the performance optimization of parallel applications on multicore processors. Therefore, all programmers have to do to benefit from SPLiT is to add a few library calls to let SPLiT know which part of the application should be analyzed. This simple but compelling optimization library contributes to enrich pervasive servers on a multicore processor, which is a strong candidate for an architecture of information appliances in the near future. SPLiT analyzes and predicts application behaviors based on CPU cycle counts and cache misses. According to the analysis and predictions, SPLiT tries to allocate processes and threads sharing data onto the same physical cores in order to enhance cache efficiency. SPLiT also tries to separate cache effective codes from the codes with more cache misses for the purpose of the avoidance of cache pollutions, which result in performance degradation. Empirical experiments assuming web applications validated the efficiency of SPLiT and the performance of the web application is improved by 26%.