Daewook Hwang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Daewook Hwang has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
- Jaeyoon Hwang
Seoul National University, Seoul, Republic of Korea
2 shared publications - Kyunghan Lee
Seoul National University, Seoul, Republic of Korea
2 shared publications - Sangtae Ha
University of Colorado Boulder, Boulder, USA
2 shared publications - Taekyoung Han
Seoul National University, Seoul, Republic of Korea
2 shared publications - Byunggu Kang
Seoul National University, Seoul, Republic of Korea
1 shared publication - Donggyu Yang
Seoul National University, Seoul, Republic of Korea
1 shared publication
Latest Publications
- Open Access
- 19 November 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Networking
- Vol. 4 (CoNEXT4) , 1-24
Unlocking Opportunistic Transmission in Datacenter Flow Control
As datacenter link bandwidth grows, hop-by-hop flow control (FC) has become essential for rapid congestion suppression by propagating backpressure to pause upstream transmission. However, since this backpressure applies at a per-queue granularity, FC suffers from Head-of-Line (HoL) blocking, where uncongested traffic is throttled unnecessarily when sharing queues with congested flows. The state-of-the-art FC schemes mitigate this by isolating congested flows into separate queues, yet such isolation is fundamentally limited by queue scarcity in commodity switches. This paper rethinks how FC reacts to congestion. Our key observation is that paused queues exhibit highly heterogeneous congestion levels, with many of them still carrying substantial uncongested traffic. Building on this insight, we present G ambit , an FC augmentation that enables opportunistic transmission under HoL blocking. By selectively relaxing pause on less congested queues based on per-queue congestion levels, G ambit allows uncongested flows to make progress by exploiting unused bandwidth, while bounding congestion risks. We verify the hardware feasibility of \gambit by prototyping it on the Intel Tofino switch with P4, and further conduct large-scale trace-driven simulations with production workloads. Our evaluation shows that G ambit effectively recovers the performance of FC lost due to practical queue constraints, reducing average FCT slowdown by up to 49.7% and 33.1%, and short-flow 99th-percentile slowdown by up to 50.6% and 55.6%, for BFC and Pyrrha, respectively.
- 19 November 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Networking
- Vol. 4 (CoNEXT4) , 1-24
Open AccessUnlocking Opportunistic Transmission in Datacenter Flow Control
As datacenter link bandwidth grows, hop-by-hop flow control (FC) has become essential for rapid congestion suppression by propagating backpressure to pause upstream transmission. However, since this backpressure applies at a per-queue granularity, FC suffers from Head-of-Line (HoL) blocking, where uncongested traffic is throttled unnecessarily when sharing queues with congested flows. The state-of-the-art FC schemes mitigate this by isolating congested flows into separate queues, yet such isolation is fundamentally limited by queue scarcity in commodity switches. This paper rethinks how FC reacts to congestion. Our key observation is that paused queues exhibit highly heterogeneous congestion levels, with many of them still carrying substantial uncongested traffic. Building on this insight, we present G ambit , an FC augmentation that enables opportunistic transmission under HoL blocking. By selectively relaxing pause on less congested queues based on per-queue congestion levels, G ambit allows uncongested flows to make progress by exploiting unused bandwidth, while bounding congestion risks. We verify the hardware feasibility of \gambit by prototyping it on the Intel Tofino switch with P4, and further conduct large-scale trace-driven simulations with production workloads. Our evaluation shows that G ambit effectively recovers the performance of FC lost due to practical queue constraints, reducing average FCT slowdown by up to 49.7% and 33.1%, and short-flow 99th-percentile slowdown by up to 50.6% and 55.6%, for BFC and Pyrrha, respectively.
- Open Access
- 1 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Networking
- Vol. 4 (CoNEXT2) , 1-23
PRC: Process-centric Rate Control for Transient Stragglers in Hybrid-Parallel Training
Stragglers remain an open challenge in large model training, forcing other GPUs to wait at synchronization barriers. While severe, persistent stragglers have been actively studied, their transient counterparts are often overlooked due to their short-lived nature. We reveal that the interplay between the transient straggler effects and network contention drastically amplifies the relative delays across workers, significantly slowing down training iterations. To address this, an intuitive approach is to prioritize bandwidth for the straggler, allowing it to catch up to the front-runners. However, while the transport layer is well-positioned to accommodate such responsive bandwidth control, conventional designs lack visibility into process-level progress. To bridge this gap, we propose PRC (Process-centric Rate Control), a new sending-rate control designed to mitigate the transient straggler effects. PRC adjusts NIC sending rates by inferring local GPU process-level information, enabling transient straggler processes to utilize more bandwidth on time. Extensive experiments on both a real-world cluster and large-scale simulations confirm that PRC effectively accelerates transient stragglers, achieving a training speedup of up to 28% compared to using state-of-the-art datacenter congestion control schemes.
- 1 June 2026
- Published by Association for Computing Machinery (ACM) in Proceedings of the ACM on Networking
- Vol. 4 (CoNEXT2) , 1-23
Open AccessPRC: Process-centric Rate Control for Transient Stragglers in Hybrid-Parallel Training
Stragglers remain an open challenge in large model training, forcing other GPUs to wait at synchronization barriers. While severe, persistent stragglers have been actively studied, their transient counterparts are often overlooked due to their short-lived nature. We reveal that the interplay between the transient straggler effects and network contention drastically amplifies the relative delays across workers, significantly slowing down training iterations. To address this, an intuitive approach is to prioritize bandwidth for the straggler, allowing it to catch up to the front-runners. However, while the transport layer is well-positioned to accommodate such responsive bandwidth control, conventional designs lack visibility into process-level progress. To bridge this gap, we propose PRC (Process-centric Rate Control), a new sending-rate control designed to mitigate the transient straggler effects. PRC adjusts NIC sending rates by inferring local GPU process-level information, enabling transient straggler processes to utilize more bandwidth on time. Extensive experiments on both a real-world cluster and large-scale simulations confirm that PRC effectively accelerates transient stragglers, achieving a training speedup of up to 28% compared to using state-of-the-art datacenter congestion control schemes.