| Project/Area Number |
17K12682
|
|---|
| Research Category |
Grant-in-Aid for Young Scientists (B)
|
| Allocation Type | Multi-year Fund |
|---|
| Research Field |
Information network
|
|---|
| Research Institution | National Institute of Informatics |
|---|
Principal Investigator |
Hu Yao 国立情報学研究所, アーキテクチャ科学研究系, 特任研究員 (50791232)
|
|---|
| Research Collaborator |
Koibuchi Michihiro
|
|---|
| Project Period (FY) |
2017-04-01 – 2019-03-31
|
| Project Status |
Completed (Fiscal Year 2018)
|
| Budget Amount *help |
¥4,030,000 (Direct Cost: ¥3,100,000、Indirect Cost: ¥930,000)
Fiscal Year 2018: ¥1,560,000 (Direct Cost: ¥1,200,000、Indirect Cost: ¥360,000)
Fiscal Year 2017: ¥2,470,000 (Direct Cost: ¥1,900,000、Indirect Cost: ¥570,000)
|
|---|
| Keywords | データセンター / 光無線 / ネットワークトポロジ / 相互結合網 / スケジューリング |
|---|
| Outline of Final Research Achievements |
We investigated the effectiveness and efficiency of an Inter-Rackscale (IRS) datacenter architecture which disaggregates hardware components such as CPU, SSD and GPU into different racks according to their own areas. By introducing FSO (free-space optics) channels for wireless connections between racks to make full use of computing resources with a fine-grained granularity during job allocation, we improved the resource utilization and communication latency for datacenter systems. We developed the methods and algorithms of job mapping and job scheduling to fully utilize the interconnection networks of racks connected by the optical wireless links. With a series of event driven simulations, we showed that the FSO links can reduce the hop count and communication latency for user jobs. We also presented the advantage of the FSO-equipped IRS systems in job scheduling performance such as average turnaround time of dispatched jobs for given sets of benchmark workloads.
|
| Academic Significance and Societal Importance of the Research Achievements |
本研究で開発したスパコンスケジューラのプログラムをオープンソースソフトウェアとして公開した。研究過程で得られた知見については、産業界・学術界の技術者・研究者らと幅広い議論を交えながら、研究会・国際会議・論文誌などで発表し、将来の光無線環境データセンターに向けた参考とする。
本研究により、低遅延光無線通信時代にあって大規模計算機システムがそのポテンシャルを十分に発揮することで、ニューラルネットワークに代表される最近のビッグデータ処理速度や並列アプリケーション実行性能をより一層向上させることが期待できる。
|
