成果報告書詳細
管理番号20160000000315
タイトル*平成27年度中間年報 エネルギー・環境新技術先導プログラム IoT時代のCPSに必要な極低消費電力データセントリック・コンピューティング技術の研究開発
公開日2016/5/17
報告書年度2015 - 2015
委託先名学校法人中央大学 株式会社東芝 株式会社PreferredNetworks
プロジェクト番号P14004
部署名イノベーション推進部
和文要約
英文要約Title: Advanced Research Program for Energy and Environmental Technologies. Extra low power data centric computing technology required for CPS in IoT era
(FY2015-FY2017) FY2015 Annual Report

For cyber physical systems (CPS) in IoT era, data centric storage system (DCSS) has been developing to realize low power, high performance, and small storage system. This research program has been developing new storage architecture with distributed management system and high-speed non-volatile memories. The DCSS processes simultaneous and multiple accesses from large amounts of sensors.

Toshiba investigated data centric storage system (DCSS) that consists of control unit (CU) and node module(NM). NM consists of a NAND flash memory and a node controller(NC). To clarify problem related with performance and power consumption of DCSS, they investigated the influence of data location for access time, and improved key-value store (KVS) database. From the investigation, data location is not major factor for access time. Most of time is spent to processing of eMMC. To improve KVS database performance and reliability, they developed data access algorithm, RAID function, and lock function.

Chuo University has developed a data management algorithm based on the characteristics of data access pattern for the storage class memory (SCM)/NAND flash hybrid memory system. The algorithm utilizes SCM as storage. As a result, the SCM/NAND flash hybrid memory system has achieved 5 times more performance. Furthermore, endurance of the NAND flash memory is extended by 4 times more.

Preferred Networks has developed a data search algorithm based on block-vector quantization methods. For images or sounds, data are often represented as real-valued vectors (deep neural networks are used for calculating these vectors). However, in general, it is difficult to search real-valued vectors efficiently. Their algorithm achieves significantly efficient search over these vectors in terms of both speed and accuracy. In simulated hardware, they confirmed their effectiveness with regard to the speed and accuracy. They are preparing several patent applications covering the proposed methods.
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