本文へジャンプ

成果報告書詳細
管理番号20190000000382
タイトル*平成30年度中間年報 高効率・高速処理を可能とするAIチップ・次世代コンピューティングの技術開発 次世代コンピューティング技術の開発 実社会の事象をリアルタイム処理可能な次世代データ処理基盤技術の研究開発
公開日2019/7/4
報告書年度2018 - 2018
委託先名日本電気株式会社 株式会社ノーチラス・テクノロジーズ 国立大学法人東京工業大学 国立大学法人大阪大学 国立大学法人名古屋大学 学校法人慶應義塾 株式会社パスコ
プロジェクト番号P16007
部署名IoT推進部
和文要約
英文要約Title: Project for Innovative AI Chips and Next-Generation Computing Technology Development/Development of Next-Generation Computing Technologies/Research Project for Next Generation Data Processing Technologies for Processing Events of Real World in Real Time (FY2018-FY2020) FY2018 Annual Report

The purpose of this project is research and development of a next generation data processing technologies, which leverage cutting edge hardware technologies. This project consists of three major parts: 1) research on core engines for fundamental data processing, 2) development of a database management system (DBMS) that encapsulates the core engines, and 3) development of applications and experiments for verification of real world use cases.

1. Core engines
(1) HTAP
Towards realizing a consistent and high-performance HTAP system, the study to guarantee the data consistency was started. A swap-out method between OLTP and OLAP that can effectively use non-volatile memory to reduce secondary storage I/Os is studied. Also, a study of an access-based query-to-CPU assignment method is performed.
(2) Streaming and 3D TIN data
We developed an approximate recovery method based on a statistical model, and proposed a task allocation algorithm to achieve effective recovery. We also developed the management model for storing 3D TINs and point clouds to relational databases and associate with their metadata flexibly.
(3) OLAP
Targeting for global/local outlier detection for OLAP queries, we developed a prototype that integrates multiple query optimization and top-k search. We evaluated the performance of SparkSQL in large-scale setting. We employ the OLAP cube data structure that effectively shares the results of multiple OLAP queries, and we incrementally maintain the OLAP cube a streaming processing manner.
(4) OLTP
We first conducted evaluation of multiple concurrency control methods proposed recently. Performance evaluation was performed in a 224 parallel environment. Second, we evaluated the integration of concurrency control and logging method. Finally, we proposed an efficient shuffling scheme and conducted evaluation on 1024 machine environment.

2. DBMS
We designed an architecture and interfaces of the database management system composed of the core engines described above. The system also includes many additional components for practical use. We developed a preliminary system that can execute the most fundamental queries for a benchmark.

3. Applications
We conducted hearings on application usages in some use cases to design our system, and preliminary experiments based on the results from the hearings, aimed at studying performance characteristics on many-core servers.
We took aerial photographs of a city by the oblique camera from airplane. We generated 3D TIN model from these photographs by existing SfM/MVS system. We modified open-source software libraries of SfM to generate 3D points cloud directly from the photographs.
ダウンロード成果報告書データベース(ユーザ登録必須)から、ダウンロードしてください。

▲トップに戻る