トップページ イベント・広報 R-CCS Cafe R-CCS Cafe 第207回 第2部
R-CCS Cafe 第207回 第2部
講演題目
HPC enhanced 1:1 scale agent based simulations of large economies for disaster-oriented applications
開催日 | 2021年2月12日(金) |
---|---|
開催時間 | 16:20 - 16:40(16:40 - 17:00 講演者を交えたフリーディスカッション(冒頭に1-2分の小休止を挟みます)、17:00 - 自由討論(参加自由)) |
開催都市 | オンライン |
場所 | BlueJeansによる遠隔セミナー |
使用言語 | 発表・スライド共に英語 |
登壇者 |
講演要旨
We pioneered developing a scalable HPC extension for Agent-Based Economic Models (ABEMs) capable of simulating 1:1 scale models of large economies with hundreds of millions of agents with the aim of disaster-oriented applications. The heavy dependency of economic entities (e.g. firms, banks, households, etc.) among each other, lifelines and other infrastructures makes the economic aftermath of localized disasters like major earthquake to cascade over the whole nation or even to the world. To ensure fast economic recovery and eliminate unforeseen long-term losses, recovery plans must be comprehensively evaluated by considering these complex inter-dependencies. However, standard economic models, like DSGE, are not capable of taking these fine-grained details into account. Although ABEMs are capable of including all these real-world complexities, lack of HPC implementations capable of simulating 1:1 scale models of major economies with hundreds of millions of agents is the major hurdle in utilizing those in disaster recovery. To overcome this hurdle, we pioneered development of a scalable HPC extension for ABEMs. In ABEMs, millions of agents interacting over several graphs, which are either centralized or scale-free in nature. While most of the interactions are bi-directional, the interaction graphs are dense, random and evolve with time. These characteristics cause a very large and unknown number of random communications among MPI processes, posing challenges to develop scalable parallel extensions. Further, random access to large volume of data makes the algorithms highly memory-bound, degrading computational performance. Adopting various strategies inspired by the real world, we drastically reduced the number of MPI communications to a known handful number, and the performance of memory bound functions are improved by implementing cache-efficient algorithms. Further, an MPI + OpenMP hybrid model is developed to best utilize modern many-core computing nodes with low per-core memory capacity, like those of Fugaku. It is demonstrated that our implementation can simulate a full-fledged economi c model with 331 million agents within 108 seconds using 128 CPU cores attaining 70% strong scalability.
注意事項
- 参加の際はPCマイクの音声・ビデオをオフにされるようお願いいたします。
- 当日の会場環境や通信状態により、やむなく配信を中止・中断する場合がございます。
- プログラムの内容、時間は予告なく変更される場合があります。
- ご使用の機器やネットワークの環境によっては、ご視聴いただけない場合がございます。
- インターネット中継に関する著作権は、主催者及び発表者に帰属します。なお、配信された映像及び音声、若しくはその内容を、理化学研究所の許可無くほかのウェブサイトや著作物等への転載,複製,改変等を行うことを禁じます。
(2021年2月8日)