Tu Tran
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Tu Tran has a total of 20 co-authors. Here are the top co-authors by number of shared publications:
- Dhabaleswar Kumar Panda
The Ohio State University, Columbus, USA
3 shared publications - Hari Subramoni
The Ohio State University, Columbus, USA
3 shared publications - Bharath Ramesh
Ohio State University, Columbus, Ohio, USA
2 shared publications - Mustafa Abduljabbar
KAUST) Mathematical and Computer Sciences and Engineering Division Thuwal King Abdullah University of Science and Technology 23955-6900 Saudi Arabia
2 shared publications Aamir Shafi
Computer Science and Engineering The Ohio State University Columbus Ohio USA
1 shared publicationBenjamin Michalowicz
Computer Science and Engineering The Ohio State University Columbus Ohio USA
1 shared publication
Latest Publications
- 8 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 45 (2) , 36-45
OHIO: Enhancing RDMA Scalability in Alltoall With Optimized Communication Overlap
The presence of exascale computers has pushed a new boundary in computing capability, posing performance challenges in parallel programming models on how to exploit such systems efficiently. The Message Passing Interface (MPI) is a dominant model for parallel programming. Among its primitives, MPI_Alltoall is a communication-intensive operation widely employed in numerous applications, yet it remains challenging to optimize. Alltoall algorithms are mainly classified into flat and hierarchical. The hierarchical designs avoid the slowdown of intra/inter-node communication by decoupling them. Hierarchical designs also reduce network congestion by limiting concurrently injected messages. This work demonstrates hierarchical designs also improve connection scalability in RDMA networks. This improvement is attributed to the cache thrashing happening inside network adapters. All of these advantages of hierarchical schemes collectively contribute to the network scalability of Alltoall. We propose and evaluate a network-agnostic design on InfiniBand and Omni-Path clusters, showing benefits at both micro-benchmark and application levels over other MPI libraries.
- 8 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Micro
- Vol. 45 (2) , 36-45
OHIO: Enhancing RDMA Scalability in Alltoall With Optimized Communication Overlap
The presence of exascale computers has pushed a new boundary in computing capability, posing performance challenges in parallel programming models on how to exploit such systems efficiently. The Message Passing Interface (MPI) is a dominant model for parallel programming. Among its primitives, MPI_Alltoall is a communication-intensive operation widely employed in numerous applications, yet it remains challenging to optimize. Alltoall algorithms are mainly classified into flat and hierarchical. The hierarchical designs avoid the slowdown of intra/inter-node communication by decoupling them. Hierarchical designs also reduce network congestion by limiting concurrently injected messages. This work demonstrates hierarchical designs also improve connection scalability in RDMA networks. This improvement is attributed to the cache thrashing happening inside network adapters. All of these advantages of hierarchical schemes collectively contribute to the network scalability of Alltoall. We propose and evaluate a network-agnostic design on InfiniBand and Omni-Path clusters, showing benefits at both micro-benchmark and application levels over other MPI libraries.
- Open Access
- 17 July 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-8
OMB-CXL: A Micro-Benchmark Suite for Evaluating MPI Communication Utilizing Compute Express Link Memory Devices
Compute Express Link (CXL) is a promising technology providing connectivity between host processors and peripheral devices like accelerators or memory modules. Compute nodes are usually connected through a high-speed network like Ethernet or Infiniband. CXL provides another way for connectivity by having compute nodes connected through CXL switches. CXL devices can connect to the switches, granting resource pooling and sharing across nodes. Access latency to CXL memory is ∼ 10x smaller than access to the memory of another node through network operations. This enables a more efficient way to communicate between nodes. Currently, no existing Message Passing Interface (MPI) libraries utilize CXL for inter-node communication. In this paper, we propose the usage of CXL memory devices to enhance message-passing communication across nodes. To demonstrate the benefit of CXL, we extend OSU Micro-Benchmark (OMB), a well-known MPI benchmark suite, to evaluate point-to-point communication going over CXL; the extended OMB is named OMB-CXL. As CXL technology is evolving and still under development, the availability of a CXL system with switches and memory devices is limited. We describe how to set up an emulated CXL system using QEMU, a virtualization software, for early experience. The experimental results show that communication over CXL is 15x better for small messages and 4x for larger ones in latency and bandwidth than over the network on average.
- 17 July 2024
- Published by Association for Computing Machinery (ACM)
- p. 1-8
Open AccessOMB-CXL: A Micro-Benchmark Suite for Evaluating MPI Communication Utilizing Compute Express Link Memory Devices
Compute Express Link (CXL) is a promising technology providing connectivity between host processors and peripheral devices like accelerators or memory modules. Compute nodes are usually connected through a high-speed network like Ethernet or Infiniband. CXL provides another way for connectivity by having compute nodes connected through CXL switches. CXL devices can connect to the switches, granting resource pooling and sharing across nodes. Access latency to CXL memory is ∼ 10x smaller than access to the memory of another node through network operations. This enables a more efficient way to communicate between nodes. Currently, no existing Message Passing Interface (MPI) libraries utilize CXL for inter-node communication. In this paper, we propose the usage of CXL memory devices to enhance message-passing communication across nodes. To demonstrate the benefit of CXL, we extend OSU Micro-Benchmark (OMB), a well-known MPI benchmark suite, to evaluate point-to-point communication going over CXL; the extended OMB is named OMB-CXL. As CXL technology is evolving and still under development, the availability of a CXL system with switches and memory devices is limited. We describe how to set up an emulated CXL system using QEMU, a virtualization software, for early experience. The experimental results show that communication over CXL is 15x better for small messages and 4x for larger ones in latency and bandwidth than over the network on average.
- 9 August 2023
- Published by Wiley in Concurrency and Computation: Practice and Experience
- Vol. 36 (1)
Accelerating communication with multi‐HCA aware collectives in MPI
Summary: To accelerate the communication between nodes, supercomputers are now equipped with multiple network adapters per node, also referred to as HCAs (Host Channel Adapters), resulting in a “multi‐rail”/“multi‐HCA” network. For example, the ThetaGPU system at Argonne National Laboratory (ANL) has eight adapters per node; with this many networking resources available, utilizing all of them becomes non‐trivial. The Message Passing Interface (MPI) is a dominant model for high‐performance computing clusters. Not all MPI collectives utilize all resources, and this becomes more apparent with advances in bandwidth and adapter count in a given cluster. In this work, we provide a thorough performance analysis of existing multirail solutions and their implications on collectives and present the necessity for further enhancement. Specifically, we propose novel designs for hierarchical, multi‐HCA‐aware Allgather. The proposed designs fully utilize all the available network adapters within a node and provide high overlap between inter‐node and intra‐node communication. At the micro‐benchmark level, we see large inter‐node improvements up to 62% and 61% better than HPC‐X and MVAPICH2‐X for 1024 processes. Because Allgather is used in Ring‐Allreduce, our designs also improve its performance by 56% and 44% compared to HPC‐X and MVAPICH2‐X, respectively. At the application level, our enhanced Allgather shows and improvement in a matrix‐vector multiplication kernel when compared to HPC‐X and MVAPICH2‐X, and Allreduce performs up to 7.83% better in deep learning training against MVAPICH2‐X.
- 9 August 2023
- Published by Wiley in Concurrency and Computation: Practice and Experience
- Vol. 36 (1)
Accelerating communication with multi‐HCA aware collectives in MPI
Summary: To accelerate the communication between nodes, supercomputers are now equipped with multiple network adapters per node, also referred to as HCAs (Host Channel Adapters), resulting in a “multi‐rail”/“multi‐HCA” network. For example, the ThetaGPU system at Argonne National Laboratory (ANL) has eight adapters per node; with this many networking resources available, utilizing all of them becomes non‐trivial. The Message Passing Interface (MPI) is a dominant model for high‐performance computing clusters. Not all MPI collectives utilize all resources, and this becomes more apparent with advances in bandwidth and adapter count in a given cluster. In this work, we provide a thorough performance analysis of existing multirail solutions and their implications on collectives and present the necessity for further enhancement. Specifically, we propose novel designs for hierarchical, multi‐HCA‐aware Allgather. The proposed designs fully utilize all the available network adapters within a node and provide high overlap between inter‐node and intra‐node communication. At the micro‐benchmark level, we see large inter‐node improvements up to 62% and 61% better than HPC‐X and MVAPICH2‐X for 1024 processes. Because Allgather is used in Ring‐Allreduce, our designs also improve its performance by 56% and 44% compared to HPC‐X and MVAPICH2‐X, respectively. At the application level, our enhanced Allgather shows and improvement in a matrix‐vector multiplication kernel when compared to HPC‐X and MVAPICH2‐X, and Allreduce performs up to 7.83% better in deep learning training against MVAPICH2‐X.
- Open Access
- 21 June 2023
- Published by Association for Computing Machinery (ACM)
- p. 477-487
Enabling Reconfigurable HPC through MPI-based Inter-FPGA Communication
- 21 June 2023
- Published by Association for Computing Machinery (ACM)
- p. 477-487
Open AccessEnabling Reconfigurable HPC through MPI-based Inter-FPGA Communication