Youngsok Kim
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
Research Subjects
Youngsok Kim has a total of 105 co-authors. Here are the top co-authors by number of shared publications:
- Jinho Lee
Seoul National University
13 shared publications - Hanjun Kim
Yonsei University, Seoul, Republic of Korea
5 shared publications - Jounghoo Lee
Yonsei University, Seoul, Republic of Korea
5 shared publications - Joonsung Kim
Sungkyunkwan University, Suwon, Republic of Korea
4 shared publications - Yeonan Ha
Yonsei University, Seoul, Republic of Korea
4 shared publications - Yongjun Park
Yonsei University, Seoul, Republic of Korea
4 shared publications
Latest Publications
- 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.
- 9 January 2026
- Published by Elsevier in Journal of Systems Architecture
- Vol. 173
Peak-memory-aware partitioning and scheduling for multi-tenant DNN model inference
- 9 January 2026
- Published by Elsevier in Journal of Systems Architecture
- Vol. 173
Peak-memory-aware partitioning and scheduling for multi-tenant DNN model inference
- Open Access
- 17 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 418-431
LATPC: Accelerating GPU Address Translation Using Locality-Aware TLB Prefetching and MSHR Compression
- 17 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 418-431
Open AccessLATPC: Accelerating GPU Address Translation Using Locality-Aware TLB Prefetching and MSHR Compression
- 23 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (2) , 915-928
DANCE++: Differentiable Accelerator/Network Co-Exploration With Hard Constraints and Data-Free Training for Real-World Scenarios
Co-exploration of neural architectures and hardware accelerators has emerged as a promising approach to address computational cost problems, especially in low-profile systems. However, existing co-exploration methods based on reinforcement learning or evolutionary search suffer from substantial search costs. To address this, this work presents DANCE++, a differentiable approach toward the co-exploration of hardware and network architecture design. At the heart of DANCE++ is a differentiable evaluator network that models hardware metrics with a neural network, enabling accelerator design through backpropagation. DANCE++ significantly reduces search time and enhances accuracy and hardware cost metrics compared to traditional approaches. To further address real-world scenarios, this work embodies two important practical topics: 1) hard constraints and 2) data dependency. To meet the constraints, such as frame rates, this work proposes a gradient manipulation algorithm that guides differentiable optimization to find hard-constrained solutions. Also to consider cases where training dataset is inaccessible, this work proposes to use data-free training methods in both co-exploration and training phases. To the best of our knowledge, DANCE++ is the first co-exploration method that targets these real-world challenges, supported by extensive experiments demonstrating its effectiveness.
- 23 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
- Vol. 45 (2) , 915-928
DANCE++: Differentiable Accelerator/Network Co-Exploration With Hard Constraints and Data-Free Training for Real-World Scenarios
Co-exploration of neural architectures and hardware accelerators has emerged as a promising approach to address computational cost problems, especially in low-profile systems. However, existing co-exploration methods based on reinforcement learning or evolutionary search suffer from substantial search costs. To address this, this work presents DANCE++, a differentiable approach toward the co-exploration of hardware and network architecture design. At the heart of DANCE++ is a differentiable evaluator network that models hardware metrics with a neural network, enabling accelerator design through backpropagation. DANCE++ significantly reduces search time and enhances accuracy and hardware cost metrics compared to traditional approaches. To further address real-world scenarios, this work embodies two important practical topics: 1) hard constraints and 2) data dependency. To meet the constraints, such as frame rates, this work proposes a gradient manipulation algorithm that guides differentiable optimization to find hard-constrained solutions. Also to consider cases where training dataset is inaccessible, this work proposes to use data-free training methods in both co-exploration and training phases. To the best of our knowledge, DANCE++ is the first co-exploration method that targets these real-world challenges, supported by extensive experiments demonstrating its effectiveness.
- Open Access
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 1509-1523
GCStack+GCScaler: Fast and Accurate GPU Performance Analyses Using Fine-Grained Stall Cycle Accounting and Interval Analysis
- 20 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 1509-1523
Open AccessGCStack+GCScaler: Fast and Accurate GPU Performance Analyses Using Fine-Grained Stall Cycle Accounting and Interval Analysis
- Open Access
- 8 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 458-472
PIM-CARE: A Compiler-Assisted Dynamic Resource Allocation Framework for Real-world DRAM PIM
- 8 June 2025
- Published by Association for Computing Machinery (ACM)
- p. 458-472
Open AccessPIM-CARE: A Compiler-Assisted Dynamic Resource Allocation Framework for Real-world DRAM PIM
- 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.
- 9 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 24 (1) , 1-4
IntervalSim++: Enhanced Interval Simulation for Unbalanced Processor Designs
As processor microarchitecture is getting complicated, an accurate analytic model becomes crucial for exploring large processor design space within limited development time. An interval simulation is a widely used analytic model for processor designs in the early stage. However, it cannot accurately model modern microarchitecture, which has an unbalanced pipeline. In this work, we introduce IntervalSim++, an accurate analytic model for a modern microarchitecture design based on the interval simulation. We identify key components highly related to the unbalanced pipeline and propose new modeling techniques atop the interval simulation without incurring significant overheads. Our evaluations show IntervalSim++ accurately models a modern out-of-order processor with minimal overheads, showing 1% average CPI error and only 8.8% simulation time increase compared to the baseline interval simulation.
- 9 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 24 (1) , 1-4
IntervalSim++: Enhanced Interval Simulation for Unbalanced Processor Designs
As processor microarchitecture is getting complicated, an accurate analytic model becomes crucial for exploring large processor design space within limited development time. An interval simulation is a widely used analytic model for processor designs in the early stage. However, it cannot accurately model modern microarchitecture, which has an unbalanced pipeline. In this work, we introduce IntervalSim++, an accurate analytic model for a modern microarchitecture design based on the interval simulation. We identify key components highly related to the unbalanced pipeline and propose new modeling techniques atop the interval simulation without incurring significant overheads. Our evaluations show IntervalSim++ accurately models a modern out-of-order processor with minimal overheads, showing 1% average CPI error and only 8.8% simulation time increase compared to the baseline interval simulation.
- Open Access
- 13 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 91-107
GraNNDis: Fast Distributed Graph Neural Network Training Framework for Multi-Server Clusters
Graph neural networks (GNNs) are one of the rapidly growing fields within deep learning. While many distributed GNN training frameworks have been proposed to increase the training throughput, they face three limitations when applied to multi-server clusters. 1) They suffer from an inter-server communication bottleneck because they do not consider the inter-/intra-server bandwidth gap, a representative characteristic of multi-server clusters. 2) Redundant memory usage and computation hinder the scalability of the distributed frameworks. 3) Sampling methods, de facto standard in mini-batch training, incur unnecessary errors in multi-server clusters. We found that these limitations can be addressed by exploiting the characteristics of multi-server clusters. Here, we propose GraNNDis, a fast distributed GNN training framework for multi-server clusters. Firstly, we present Flexible Preloading, which preloads the essential vertex dependencies server-wise to reduce the low-bandwidth inter-server communications. Secondly, we introduce Cooperative Batching, which enables memory-efficient, less redundant mini-batch training by utilizing high-bandwidth intra-server communications. Thirdly, we propose Expansion-aware Sampling, a cluster-aware sampling method, which samples the edges that affect the system speedup. As sampling the intra-server dependencies does not contribute much to the speedup as they are communicated through fast intra-server links, it only targets a server boundary to be sampled. Lastly, we introduce One-Hop Graph Masking, a computation and communication structure to realize the above methods in multi-server environments. We evaluated GraNNDis on multi-server clusters, and it provided significant speedup over the state-of-the-art distributed GNN training frameworks. GraNNDis is open-sourced at https://github.com/AIS-SNU/GraNNDis_Artifact to facilitate its use.
- 13 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 91-107
Open AccessGraNNDis: Fast Distributed Graph Neural Network Training Framework for Multi-Server Clusters
Graph neural networks (GNNs) are one of the rapidly growing fields within deep learning. While many distributed GNN training frameworks have been proposed to increase the training throughput, they face three limitations when applied to multi-server clusters. 1) They suffer from an inter-server communication bottleneck because they do not consider the inter-/intra-server bandwidth gap, a representative characteristic of multi-server clusters. 2) Redundant memory usage and computation hinder the scalability of the distributed frameworks. 3) Sampling methods, de facto standard in mini-batch training, incur unnecessary errors in multi-server clusters. We found that these limitations can be addressed by exploiting the characteristics of multi-server clusters. Here, we propose GraNNDis, a fast distributed GNN training framework for multi-server clusters. Firstly, we present Flexible Preloading, which preloads the essential vertex dependencies server-wise to reduce the low-bandwidth inter-server communications. Secondly, we introduce Cooperative Batching, which enables memory-efficient, less redundant mini-batch training by utilizing high-bandwidth intra-server communications. Thirdly, we propose Expansion-aware Sampling, a cluster-aware sampling method, which samples the edges that affect the system speedup. As sampling the intra-server dependencies does not contribute much to the speedup as they are communicated through fast intra-server links, it only targets a server boundary to be sampled. Lastly, we introduce One-Hop Graph Masking, a computation and communication structure to realize the above methods in multi-server environments. We evaluated GraNNDis on multi-server clusters, and it provided significant speedup over the state-of-the-art distributed GNN training frameworks. GraNNDis is open-sourced at https://github.com/AIS-SNU/GraNNDis_Artifact to facilitate its use.
- 9 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 23 (2) , 235-238
GCStack: A GPU Cycle Accounting Mechanism for Providing Accurate Insight Into GPU Performance
Cycles Per Instruction (CPI) stacks help computer architects gain insight into the performance of their target architectures and applications. To bring the benefits of CPI stacks to Graphics Processing Units (GPUs), prior studies have proposed GPU cycle accounting mechanisms that can identify the stall cycles and their stall events on GPU architectures. Unfortunately, the prior studies cannot provide accurate insight into the GPU performance due to their coarse-grained, priority-driven, and issue-centric cycle accounting mechanisms. In this paper, we present GCStack , a fine-grained GPU cycle accounting mechanism that constructs accurate CPI stacks and accurately identifies primary GPU performance bottlenecks. GCStack first exposes all the stall events of the outstanding warps of a warp scheduler, most of which get hidden by the existing mechanisms. Then, GCStack defers the classification of structural stalls, which the existing mechanisms cannot correctly identify with their issue-stage-centric stall classification, to the later stages of the GPU pipeline. We implement GCStack on Accel-Sim and show that GCStack provides more accurate CPI stacks and GPU performance insight than GSI, the state-of-the-art GPU cycle accounting mechanism whose primary focus is on characterizing memory-related stalls.
- 9 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Computer Architecture Letters
- Vol. 23 (2) , 235-238
GCStack: A GPU Cycle Accounting Mechanism for Providing Accurate Insight Into GPU Performance
Cycles Per Instruction (CPI) stacks help computer architects gain insight into the performance of their target architectures and applications. To bring the benefits of CPI stacks to Graphics Processing Units (GPUs), prior studies have proposed GPU cycle accounting mechanisms that can identify the stall cycles and their stall events on GPU architectures. Unfortunately, the prior studies cannot provide accurate insight into the GPU performance due to their coarse-grained, priority-driven, and issue-centric cycle accounting mechanisms. In this paper, we present GCStack , a fine-grained GPU cycle accounting mechanism that constructs accurate CPI stacks and accurately identifies primary GPU performance bottlenecks. GCStack first exposes all the stall events of the outstanding warps of a warp scheduler, most of which get hidden by the existing mechanisms. Then, GCStack defers the classification of structural stalls, which the existing mechanisms cannot correctly identify with their issue-stage-centric stall classification, to the later stages of the GPU pipeline. We implement GCStack on Accel-Sim and show that GCStack provides more accurate CPI stacks and GPU performance insight than GSI, the state-of-the-art GPU cycle accounting mechanism whose primary focus is on characterizing memory-related stalls.