Realm: a fully asynchronous event-based runtime system, the low-level runtime of Legion• Implemented a Realm module to offload compute from Legion applications to Xilinx FPGAs using the OpenCL extension of the Xilinx Runtime Library (XRT).• Added specialized channels in the Realm DMA engine where Legion can launch user-defined FPGA kernels.• Tested the functionality of the Realm module with multiple kernels (vector addiction, image processing and data compression).Flexflow: a distributed deep learning framework that accelerates distributed DNN training by automatically discovering fast parallelization strategies. • Improved the machine model in the Flexflow simulator, especially the communication modelling.• Designed a user-friendly and flexible configure file to describe the hardware topology of the machines.• Validated the new machine model with Legion applications (Circuit, Stencil) and ML models (Alexnet, DLRM, Resent), and all the estimations made with the new model of the tested applications are within 10% of the actual performance.

Design and Implement collective operation frameworks in Open MPI.ADAPT – an event-based collective operation framework in Open MPI. 2014-2017 • Relaxed the synchronizations to alleviate noise propagations.• Improved the performance of collective operations with hardware topology awareness.• Extended collective operations to support GPU data.• Published in HPDC 2018.HAN – a hierarchical autotuned collective framework in Open MPI. 2018-2020• Developed a task-based mechanism to combine and overlap the collective operations from different hardware levels.• Designed a novel cost model to predict the performance of collective operations. • Created an autotuning component by utilizing the new cost model to reduce the time of autotuning.• Benchmarked this framework on 4K cores, achieving significant speedup on MPI_Bcast and MPI_Allreduce, comparing to the default framework in Open MPI.• Published in Cluster 2020 (best paper award).SOLO – an intra-node collective framework in Open MPI. 2019-2020• Designing a lock-free memory pool to manage the shared memory among processes on a node.• Utilizing the shared memory space to reduce the number of memory copies.

Build in-situ data analysis infrastructure for an earth system simulator.• Automated code instrumentation of an earth system simulator to extract data of interest.• Efficiently packed and transferred customized data types.• Integrated with external data processing and machine learning packages.