Jinsun Yoo
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Jinsun Yoo has a total of 15 co-authors. Here are the top co-authors by number of shared publications:
- Alexandros Daglis
EcoCloud, EPFL
1 shared publication - Anirudh Sarma
College of Computing, Georgia Tech, Atlanta, Georgia, USA
1 shared publication - Benjamin Klenk
NVIDIA, Santa Clara, CA, USA
1 shared publication - Divya Kiran Kadiyala
Future Technologies (SIAD), Hewlett Packard Enterprise, Milpitas, California, USA
1 shared publication - Jithin Kallukalam Sojan
Georgia Institute of Technology, Atlanta, USA
1 shared publication - Lianjie Cao
Networking and Distributed Systems Lab (NDSL), Hewlett Packard Enterprise Labs, Milpitas, California, USA
1 shared publication
Latest Publications
- Open Access
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 2102-2108
Beyond Monoliths: Enabling Flexible and Composable AI Systems via Memory Disaggregation
- 11 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 2102-2108
Open AccessBeyond Monoliths: Enabling Flexible and Composable AI Systems via Memory Disaggregation
- 5 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 45 (2) , 46-55
Toward a Standardized Representation for Deep Learning Collective Algorithms
The explosion of machine learning model size has led to its execution on distributed clusters at a large scale. Many works have tried to optimize the process of producing collective algorithms and running collective communications, which act as a bottleneck to distributed machine learning. However, different works use their own collective algorithm representation, resulting in a fragmented optimization environment. We propose a standardized workflow leveraging a common representation format based on Chakra Execution Trace, a widely used graph-based representation of distributed machine learning workloads. Such a common representation enables us to view collective communications at the same level as workload operations and decouple producer and consumer tools, enhancing interoperability and reducing engineering effort. We provide a proof-of-concept of this standardized workflow by simulating collective algorithms generated by the MSCCLang domain-specific language and TACOS synthesizer through the ASTRA-sim distributed machine learning simulator using multiple network configurations.
- 5 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 45 (2) , 46-55
Toward a Standardized Representation for Deep Learning Collective Algorithms
The explosion of machine learning model size has led to its execution on distributed clusters at a large scale. Many works have tried to optimize the process of producing collective algorithms and running collective communications, which act as a bottleneck to distributed machine learning. However, different works use their own collective algorithm representation, resulting in a fragmented optimization environment. We propose a standardized workflow leveraging a common representation format based on Chakra Execution Trace, a widely used graph-based representation of distributed machine learning workloads. Such a common representation enables us to view collective communications at the same level as workload operations and decouple producer and consumer tools, enhancing interoperability and reducing engineering effort. We provide a proof-of-concept of this standardized workflow by simulating collective algorithms generated by the MSCCLang domain-specific language and TACOS synthesizer through the ASTRA-sim distributed machine learning simulator using multiple network configurations.
- Open Access
- 24 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 78-89
FEO: Efficient Resource Allocation for FaaS at the Edge
Geo-distributed Edge sites are expected to cater to the stringent demands of situation-aware applications like collaborative autonomous vehicles and drone swarms. While clients of such applications benefit from having network-proximal compute resources, an Edge site has limited resources compared to the traditional Cloud. Moreover, the load experienced by an Edge site depends on a client's mobility pattern, which may often be unpredictable. The Function-as-a-Service (FaaS) paradigm is poised aptly to handle the ephemeral nature of workload demand at Edge sites. In FaaS, applications are decomposed into containerized functions enabling fine-grained resource management. However, spatio-temporal variations in client mobility can still lead to rapid saturation of resources beyond the capacity of an Edge site. To address this challenge, we develop FEO (Federated Edge Orchestrator), a resource allocation scheme across the geodistributed Edge infrastructure for FaaS. FEO employs a novel federated policy to offload function invocations to peer sites with spare resource capacity without the need to frequently share knowledge about available capacities among participating sites. Detailed experiments show that FEO's approach can reduce a site's P99 latency by almost 3x, while maintaining application service level objectives at all other sites.
- 24 June 2024
- Published by Association for Computing Machinery (ACM)
- p. 78-89
Open AccessFEO: Efficient Resource Allocation for FaaS at the Edge
Geo-distributed Edge sites are expected to cater to the stringent demands of situation-aware applications like collaborative autonomous vehicles and drone swarms. While clients of such applications benefit from having network-proximal compute resources, an Edge site has limited resources compared to the traditional Cloud. Moreover, the load experienced by an Edge site depends on a client's mobility pattern, which may often be unpredictable. The Function-as-a-Service (FaaS) paradigm is poised aptly to handle the ephemeral nature of workload demand at Edge sites. In FaaS, applications are decomposed into containerized functions enabling fine-grained resource management. However, spatio-temporal variations in client mobility can still lead to rapid saturation of resources beyond the capacity of an Edge site. To address this challenge, we develop FEO (Federated Edge Orchestrator), a resource allocation scheme across the geodistributed Edge infrastructure for FaaS. FEO employs a novel federated policy to offload function invocations to peer sites with spare resource capacity without the need to frequently share knowledge about available capacities among participating sites. Detailed experiments show that FEO's approach can reduce a site's P99 latency by almost 3x, while maintaining application service level objectives at all other sites.