시계열 운행기록 데이터의 3계층 그래프 모델 설계 및 적재 성능 비교 연구 = Three-Tier Graph Model Design and Ingestion Performance Comparison for Time-Series Vehicle Operation Data

다국어 초록 (Multilingual Abstract)

In many transportation data platforms, road network information is stored in relational or spatial databases, while time-series vehicle operation records are managed separately in distributed stores such as HBase. In this separated architecture, queries that trace trips through specific nodes or links must repeatedly perform joins or additional mapping, and the associated overhead becomes a major bottleneck as data volume grows. To alleviate these limitations, this study designs a three-tier graph data model (node–link–record) that represents the national standard node–link road network and time-series operation logs within a single graph, and empirically evaluates its practicality from the perspective of large-scale data ingestion.
The proposed model organizes road nodes, road links, and vehicle operation records as separate vertex tiers and directly connects inter-tier continuity using edges, enabling path-based exploration without relational joins. It is implemented on the distributed graph database JanusGraph, where real vehicle operation logs are transformed into graph structures through an integrated pipeline of preprocessing, map matching, and storage. On this basis, the study quantitatively analyzes how internal index configuration, insertion-related parameter tuning, multi-threading, and backend storage selection (HBase, Cassandra, ScyllaDB) affect ingestion throughput and latency.
Experimental results show that, without even minimal internal composite indexes on key identifier attributes used for inter-tier connections, repeated full scans during edge construction make ingestion practically infeasible, whereas indexing these essential ID attributes enables stable insertion. Tuning core parameter combinations, including batch loading, yields throughput improvements of several times over the default settings, and applying multi-threading achieves approximately 1.5× additional speedup compared to single-threaded insertion. In backend storage comparisons, ScyllaDB exhibits clearly higher ingestion throughput than HBase and Cassandra, particularly in the record tier. These findings indicate that the proposed three-tier graph model is practically applicable to large-scale time-series vehicle operation data, and that index design, internal parameter combinations, and backend storage choice jointly act as key design knobs governing ingestion performance.

목차 (Table of Contents)

• Ⅰ. 서론 1
• Ⅱ. 배경 및 관련 연구 4
• 2.1 데이터 저장 및 처리 기술 4
• 2.2 JanusGraph의 필요성 5
• 2.3 JanusGraph 아키텍처 8
• 2.4 기존 연구 동향 12
• 2.5 본 연구의 필요성 14
• Ⅲ. 설계 및 구현 15
• 3.1 그래프 데이터 처리 아키텍처 15
• 3.2 그래프 데이터베이스 데이터 모델 설계 19
• 3.3 인덱스 설계 및 설정 20
• 3.4 삽입 성능 최적화를 위한 파라미터 설정 21
• 3.5 그래프 데이터베이스 데이터 적재 방안 22
• Ⅳ. 실험 26
• 4.1 실험 환경 26
• 4.2 실험 과정 29
• 4.3 실험 결과 31
• 4.3.1 파라미터 설정 실험 결과 31
• 4.3.2 멀티스레드 실험 결과 33
• 4.3.3 백엔드 스토리지 실험 결과 35
• Ⅴ. 결론 39
• 참고문헌 41