SYNCOPADE: Asynchronous Distributed Execution Substrate for Iterative Evaluation Workloads

Mitsuru Endo

2026 IEEE International Conference on Systems, Man, and Cybernetics (SMC2026)

講演番号：WeB7.4

• 講演情報 / Presentation details

---

背景 – Background

高精度なシミュレーションを繰り返す設計最適化では，評価1回あたりの計算負荷と評価回数の増大が総実行時間を押し上げる．本研究では，組織内の利用可能な異種ノードを活用し，最適化・解析が生成する評価タスクを分散実行する基盤SYNCOPADEを提案する．評価モデルの精度や問題設定を維持したまま，計算資源を有効に使うことを狙う．

Engineering design optimization involving repeated high-fidelity simulations faces increasing computational demand from both the cost and number of evaluations. This study presents SYNCOPADE, an execution substrate that distributes evaluation tasks generated by optimization or analysis across heterogeneous organizational nodes. It aims to use available computational resources while preserving model fidelity and the original problem requirements.

---

手法 – Method

SYNCOPADEは，評価点の生成と結果の統合を行うCoordinator，タスクキューとWorkerの状態を管理するConductor，評価を実行するWorkerの3役から構成される．Conductorは，実行可能なWorkerが確定した時点で待機中のタスクを割り当てる．計算結果はWorkerからCoordinatorへ直接返し，完了通知などの実行状態はConductorで管理することで，結果と制御の流れを分離する．

SYNCOPADE separates three roles: the Coordinator generates evaluation points and integrates results, the Conductor manages the task queue and Worker availability, and Workers execute evaluations. Tasks are assigned through late-binding dispatch when a Worker becomes available. Computed results return directly from Workers to the Coordinator, while completion notifications and execution state are handled by the Conductor, separating result flow from control flow.

Fig. 1 Basic three-role architecture of SYNCOPADE

階層的な評価では，Workerが一時的にsub-Coordinatorとなり，下位の評価タスクを同じConductorへ再投入する．これにより，入れ子の最適化や複合領域設計最適化（MDO）を，同一のWorkerプールを使って実行できる．

For hierarchical workloads, a Worker temporarily acts as a sub-Coordinator and resubmits lower-level evaluation tasks to the same Conductor. This allows nested optimization and multidisciplinary design optimization (MDO) workflows to reuse the same Worker pool across hierarchy levels.

Fig. 2 Extended architecture with sub-Coordinators for hierarchical workloads

---

結果 – Result

検証実験とハイパーパラメータ評価・階層型CO-MDOの事例により，評価タスクが分散実行のオーバーヘッドを吸収できる粒度では，実行時間の短縮が確認された．また，階層的なタスクの再投入も同じ基盤上で実行でき，その効率には上位の並列度と下位のバッチサイズのバランスが重要であることが示された．

Validation studies and application cases involving hyperparameter evaluation and hierarchical CO-MDO showed reduced wall-clock time when evaluation tasks were sufficiently coarse to amortize distributed overhead. Hierarchical task resubmission also operated on the same substrate, with efficiency depending on the balance between upper-level parallelism and lower-level batch size.

---

結論 – Conclusion

SYNCOPADEにより，評価点の生成・結果の統合とノード単位の実行管理を分離し，組織内の計算資源を使う非同期分散実行を構成した．今後は，より大きなWorkerプールでの評価やConductorの負荷分析，並列度・バッチサイズなどの実行パラメータの適応的な調整を進める．

SYNCOPADE separates evaluation-point generation and result integration from Worker-level execution management, providing asynchronous distributed execution using organizationally available nodes. Future work will evaluate larger Worker pools, analyze Conductor load, and investigate adaptive control of execution parameters such as parallelism and batch size.

---

This work was supported by JSPS KAKENHI Grant Number 23K03755.

---

The content on this page is not freely available for reproduction or redistribution. Unauthorized use may lead to legal consequences.