Kwanghyun Park
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Kwanghyun Park has a total of 32 co-authors. Here are the top co-authors by number of shared publications:
- Hyeonjun An
Yonsei University, Seoul, Seoul, Republic of Korea
2 shared publications - Hyunjoon Kim
Yonsei University, Seoul, Seoul, Republic of Korea
2 shared publications - Rathijit Sen
Microsoft, Redmond, WA, USA
2 shared publications - Youngsok Kim
Yonsei University, Seoul, Republic of Korea
2 shared publications - Chaemin Lim
Yonsei University, Seoul, South Korea
1 shared publication - Chaerim Lim
BDAI Lab, Yonsei University, Seoul, Republic of Korea
1 shared publication
Latest Publications
- Open Access
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
CANNON: A CXL-Based Near-Memory Processing Architecture for Approximate Nearest Neighbor Search on Real Hardware
- 26 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-7
Open AccessCANNON: A CXL-Based Near-Memory Processing Architecture for Approximate Nearest Neighbor Search on Real Hardware
- Open Access
- 18 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (3) , 1-29
DFLOP: A Data-driven Framework for Multimodal LLM Training Pipeline Optimization
Multimodal Large Language Models (MLLMs) have achieved remarkable advances by integrating text, image, and audio understanding within a unified architecture. However, existing distributed training frameworks remain fundamentally data-blind: they parallelize computation without accounting for variations in input data characteristics. This data unawareness leads to severe computation skew across stages and microbatches, where heterogeneous multimodal inputs incur different processing costs. Consequently, GPU resources are unevenly utilized, synchronization delays accumulate, and overall training efficiency degrades. To address this limitation, we present DFLOP, a d ata-driven f ramework for multimodal L LM training pipeline op timization. DFLOP continuously profiles runtime behavior to capture data-induced computation variance and employs predictive scheduling to balance workloads across stages and microbatches. By coupling data characteristics with execution planning, DFLOP substantially improves GPU utilization and throughput. Extensive experiments on large-scale multimodal benchmarks show that DFLOP achieves up to 3.6× faster training compared to state-of-the-art distributed training frameworks.
- 18 May 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 4 (3) , 1-29
Open AccessDFLOP: A Data-driven Framework for Multimodal LLM Training Pipeline Optimization
Multimodal Large Language Models (MLLMs) have achieved remarkable advances by integrating text, image, and audio understanding within a unified architecture. However, existing distributed training frameworks remain fundamentally data-blind: they parallelize computation without accounting for variations in input data characteristics. This data unawareness leads to severe computation skew across stages and microbatches, where heterogeneous multimodal inputs incur different processing costs. Consequently, GPU resources are unevenly utilized, synchronization delays accumulate, and overall training efficiency degrades. To address this limitation, we present DFLOP, a d ata-driven f ramework for multimodal L LM training pipeline op timization. DFLOP continuously profiles runtime behavior to capture data-induced computation variance and employs predictive scheduling to balance workloads across stages and microbatches. By coupling data characteristics with execution planning, DFLOP substantially improves GPU utilization and throughput. Extensive experiments on large-scale multimodal benchmarks show that DFLOP achieves up to 3.6× faster training compared to state-of-the-art distributed training frameworks.
- Open Access
- 18 December 2024
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 2 (6) , 1-27
SPID-Join: A Skew-resistant Processing-in-DIMM Join Algorithm Exploiting the Bank- and Rank-level Parallelisms of DIMMs
Recent advances in Dual In-line Memory Modules (DIMMs) allow DIMMs to support Processing-In-DIMM (PID) by placing In-DIMM Processors (IDPs) near their memory banks. Prior studies have shown that in-memory joins can benefit from PID by offloading their operations onto the IDPs and exploiting the high internal memory bandwidth of DIMMs. Aimed at evenly balancing the computational loads between the IDPs, the existing algorithms perform IDP-wise global partitioning on input tables and then make each IDP process a partition of the input tables. Unfortunately, we find that the existing PID join algorithms achieve low performance and scalability with skewed input tables. With skewed input tables, the IDP-wise global partitioning incurs imbalanced loads between the IDPs, making the IDPs remain idle until the heaviest-load IDP completes processing its partition. To fully exploit the IDPs for accelerating in-memory joins involving skewed input tables, therefore, we need a new PID join algorithm which achieves high skew resistance by mitigating the imbalanced inter-IDP loads. In this paper, we present SPID-Join, a skew-resistant PID join algorithm which exploits two parallelisms inherent in DIMM architectures, namely bank- and rank-level parallelisms. By replicating join keys across the banks within a rank and across ranks, SPID-Join significantly increases the internal memory bandwidth and computational throughput allocated to each join key, improving the load balance between the IDPs and accelerating join executions. SPID-Join exploits the bank- and the rank-level parallelisms to minimize join key replication overheads and support a wider range of join key replication ratios. Despite achieving high skew resistance, SPID-Join exhibits a trade-off between the join key replication ratio and the join execution latency, making the best-performing join key replication ratio depend on join and PID system configurations. We, therefore, augment SPID-Join with a cost model which identifies the best-performing join key replication ratio for given join and PID system configurations. By accurately modeling and scaling the IDPs' throughput and the inter-IDP communication bandwidth, the cost model accurately captures the impact of the join key replication ratio on SPID-Join. Our experimental results using eight UPMEM DIMMs, which collectively provide a total of 1,024 IDPs, show that SPID-Join achieves up to 10.38x faster join executions over PID-Join, the state-of-the-art PID join algorithm, with highly skewed input tables.
- 18 December 2024
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Management of Data
- Vol. 2 (6) , 1-27
Open AccessSPID-Join: A Skew-resistant Processing-in-DIMM Join Algorithm Exploiting the Bank- and Rank-level Parallelisms of DIMMs
Recent advances in Dual In-line Memory Modules (DIMMs) allow DIMMs to support Processing-In-DIMM (PID) by placing In-DIMM Processors (IDPs) near their memory banks. Prior studies have shown that in-memory joins can benefit from PID by offloading their operations onto the IDPs and exploiting the high internal memory bandwidth of DIMMs. Aimed at evenly balancing the computational loads between the IDPs, the existing algorithms perform IDP-wise global partitioning on input tables and then make each IDP process a partition of the input tables. Unfortunately, we find that the existing PID join algorithms achieve low performance and scalability with skewed input tables. With skewed input tables, the IDP-wise global partitioning incurs imbalanced loads between the IDPs, making the IDPs remain idle until the heaviest-load IDP completes processing its partition. To fully exploit the IDPs for accelerating in-memory joins involving skewed input tables, therefore, we need a new PID join algorithm which achieves high skew resistance by mitigating the imbalanced inter-IDP loads. In this paper, we present SPID-Join, a skew-resistant PID join algorithm which exploits two parallelisms inherent in DIMM architectures, namely bank- and rank-level parallelisms. By replicating join keys across the banks within a rank and across ranks, SPID-Join significantly increases the internal memory bandwidth and computational throughput allocated to each join key, improving the load balance between the IDPs and accelerating join executions. SPID-Join exploits the bank- and the rank-level parallelisms to minimize join key replication overheads and support a wider range of join key replication ratios. Despite achieving high skew resistance, SPID-Join exhibits a trade-off between the join key replication ratio and the join execution latency, making the best-performing join key replication ratio depend on join and PID system configurations. We, therefore, augment SPID-Join with a cost model which identifies the best-performing join key replication ratio for given join and PID system configurations. By accurately modeling and scaling the IDPs' throughput and the inter-IDP communication bandwidth, the cost model accurately captures the impact of the join key replication ratio on SPID-Join. Our experimental results using eight UPMEM DIMMs, which collectively provide a total of 1,024 IDPs, show that SPID-Join achieves up to 10.38x faster join executions over PID-Join, the state-of-the-art PID join algorithm, with highly skewed input tables.
- Open Access
- 28 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 11945-11962
Trinity: In-Database Near-Data Machine Learning Acceleration Platform for Advanced Data Analytics
The ability to perform machine learning (ML) tasks in a database management system (DBMS) is a new paradigm for conventional database systems as it enables advanced data analytics on top of well-established capabilities of DBMSs. However, the integration of ML in DBMSs introduces new challenges in traditional CPU-based systems because of its higher computational demands and bigger data bandwidth requirements. To address this, hardware acceleration has become even more important in database systems, and the computational storage device (CSD) placing an accelerator near storage is considered as an effective solution due to its high processing power with no extra data movement cost. In this paper, we propose Trinity, an end-to-end database system that enables in-database, in-storage platform that accelerates advanced analytics queries invoking trained ML models along with complex data operations. By designing a full stack from DBMS’s internal software components to hardware accelerator, Trinity enables in-database ML pipelines on the CSD. On the software side, we extend the internals of conventional DBMSs to utilize the accelerator in the SmartSSD. Our extended analyzer evaluates the compatibility of the current query with our hardware accelerator and compresses compatible queries into a 24-byte numeric format for efficient hardware processing. Furthermore, the predictor is extended to integrate our performance cost models to always offload queries into the optimal hardware backend. The proposed SmartSSD cost model mathematically models our hardware, including host operations, data transfers, FPGA kernel execution time, and the CPU cost model uses polynomial regression ML models to predict complex CPU latency. On the hardware side, we introduce the in-database processing accelerator (i-DPA), a custom FPGA-based accelerator. i-DPA includes database page decoder to fully exploit the bandwidth benefit of near-storage processing. It also employs dynamic tuple binding to enhance the overall parallelism and hardware utilization. i-DPA’s architecture having heterogeneous computing units with a reconfigurable on-chip interconnect also allows seamless data streaming, enabling task-level pipeline across different computing units. Finally, our evaluation shows that Trinity improves the end-to-end performance of analytics queries by 15.21× on average and up to 57.18× compared to the conventional CPU-based DBMS platform. We also show that the Trinity’s performance can linearly scale up with multiple SmartSSDs, achieving nearly up to 200× speedup over the baseline with four SmartSSDs.
- 28 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 11945-11962
Open AccessTrinity: In-Database Near-Data Machine Learning Acceleration Platform for Advanced Data Analytics
The ability to perform machine learning (ML) tasks in a database management system (DBMS) is a new paradigm for conventional database systems as it enables advanced data analytics on top of well-established capabilities of DBMSs. However, the integration of ML in DBMSs introduces new challenges in traditional CPU-based systems because of its higher computational demands and bigger data bandwidth requirements. To address this, hardware acceleration has become even more important in database systems, and the computational storage device (CSD) placing an accelerator near storage is considered as an effective solution due to its high processing power with no extra data movement cost. In this paper, we propose Trinity, an end-to-end database system that enables in-database, in-storage platform that accelerates advanced analytics queries invoking trained ML models along with complex data operations. By designing a full stack from DBMS’s internal software components to hardware accelerator, Trinity enables in-database ML pipelines on the CSD. On the software side, we extend the internals of conventional DBMSs to utilize the accelerator in the SmartSSD. Our extended analyzer evaluates the compatibility of the current query with our hardware accelerator and compresses compatible queries into a 24-byte numeric format for efficient hardware processing. Furthermore, the predictor is extended to integrate our performance cost models to always offload queries into the optimal hardware backend. The proposed SmartSSD cost model mathematically models our hardware, including host operations, data transfers, FPGA kernel execution time, and the CPU cost model uses polynomial regression ML models to predict complex CPU latency. On the hardware side, we introduce the in-database processing accelerator (i-DPA), a custom FPGA-based accelerator. i-DPA includes database page decoder to fully exploit the bandwidth benefit of near-storage processing. It also employs dynamic tuple binding to enhance the overall parallelism and hardware utilization. i-DPA’s architecture having heterogeneous computing units with a reconfigurable on-chip interconnect also allows seamless data streaming, enabling task-level pipeline across different computing units. Finally, our evaluation shows that Trinity improves the end-to-end performance of analytics queries by 15.21× on average and up to 57.18× compared to the conventional CPU-based DBMS platform. We also show that the Trinity’s performance can linearly scale up with multiple SmartSSDs, achieving nearly up to 200× speedup over the baseline with four SmartSSDs.
- Open Access
- 24 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 35 (10) , 10295-10308
Pushing ML Predictions Into DBMSs
In the past decade, many approaches have been suggested to execute ML workloads on a DBMS. However, most of them have looked at in-DBMS ML from a training perspective, whereas ML inference has been largely overlooked. We think that this is an important gap to fill for two main reasons: (1) in the near future, every application will be infused with some sort of ML capability; (2) behind every web page, application, and enterprise there is a DBMS, whereby in-DBMS inference is an appealing solution both for efficiency (e.g., less data movement), performance (e.g., cross-optimizations between relational operators and ML) and governance. In this paper, we study whether DBMSs are a good fit for prediction serving. We introduce a technique for translating trained ML pipelines containing both featurizers (e.g., one-hot encoding) and models (e.g., linear and tree-based models) into SQL queries, and we compare in-DBMS performance against popular ML frameworks such as Sklearn and ml.net . Our experiments show that, when pushed inside a DBMS, trained ML pipelines can have performance comparable to ML frameworks in several scenarios, while they perform quite poorly on text featurization and over (even simple) neural networks.
- 24 April 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 35 (10) , 10295-10308
Open AccessPushing ML Predictions Into DBMSs
In the past decade, many approaches have been suggested to execute ML workloads on a DBMS. However, most of them have looked at in-DBMS ML from a training perspective, whereas ML inference has been largely overlooked. We think that this is an important gap to fill for two main reasons: (1) in the near future, every application will be infused with some sort of ML capability; (2) behind every web page, application, and enterprise there is a DBMS, whereby in-DBMS inference is an appealing solution both for efficiency (e.g., less data movement), performance (e.g., cross-optimizations between relational operators and ML) and governance. In this paper, we study whether DBMSs are a good fit for prediction serving. We introduce a technique for translating trained ML pipelines containing both featurizers (e.g., one-hot encoding) and models (e.g., linear and tree-based models) into SQL queries, and we compare in-DBMS performance against popular ML frameworks such as Sklearn and ml.net . Our experiments show that, when pushed inside a DBMS, trained ML pipelines can have performance comparable to ML frameworks in several scenarios, while they perform quite poorly on text featurization and over (even simple) neural networks.
- Open Access
- 1 August 2019
- Published by Institute of Foreign Language Education in Studies in Foreign Language Education
- Vol. 33 (3) , 1-26
A study of university English learners’ oral presentation with multi-dimensional analysis: Focusing on informative and persuasive presentation
- 1 August 2019
- Published by Institute of Foreign Language Education in Studies in Foreign Language Education
- Vol. 33 (3) , 1-26
Open AccessA study of university English learners’ oral presentation with multi-dimensional analysis: Focusing on informative and persuasive presentation
- Open Access
- 1 January 2015
- Published by Architectural Institute of Japan in Journal of Architecture and Planning (Transactions of AIJ)
- Vol. 80 (713) , 1689-1696
ON THE CHANGE OF GREEN AREA IN THE ZONING OF SEOUL CITY PLANNING
- 1 January 2015
- Published by Architectural Institute of Japan in Journal of Architecture and Planning (Transactions of AIJ)
- Vol. 80 (713) , 1689-1696
Open AccessON THE CHANGE OF GREEN AREA IN THE ZONING OF SEOUL CITY PLANNING