Nicolas Castet Email & Phone Number

Santa Clara, Ca, Us

Tel Aviv, Center, Il

Distributed deep learning training at scale.

Armonk, New York, Ny, Us

• Contribute to deliver a deep learning framework suite including TensorFlow, PyTorch, NCCL and Horovod for IBM Power servers• Collaborate as a member of the Technical Steering Committee to the open source project Horovod• Tune models and frameworks to scale to thousands of GPUs on the Summit supercomputer• Streamline the deployment of distributed deep learning jobs using Kubernetes and OpenShift• Lead the effort for an elastic distributed deep learning training solution to dynamically scale up and down GPUs• Optimize the ResNet50 model to achieve state-of-the-art performance and convergence with hundreds of GPUs• Develop a multidimensional ring-based GPU communication library (IBM PowerAI DDL)• Integrated our efficient communication library with TensorFlow distribution strategy API

Armonk, New York, Ny, Us

• Contributed to the core development of NovaLink layer to bring IBM servers to cloud computing platforms OpenStack and PowerVC• Led NovaLink development to manage PowerVM and KVM hypervisors from the same platform• Optimized VM live migration flow to reduce host evacuation time• Developed VM live migration capability for Linux IO storage and SDN environment• Implemented the Elastic Stack to collect, ship, filter, and analyze virtualization logs in the cloud

• Lead the development of a new product TransportPlanner™ and its implementationfor our client NATO • Develop algorithms to optimize the transport of products in a pipeline network using graph algorithms and linear and mixed integer programming• Research various techniques to solve pipeline and terminal scheduling problems• Involved in all the stages of the software life cycle from requirement gathering to deployment and support• Developed a multithreaded tool to run efficiently pipeline optimization simulationsand browse their results

College Station, Tx, Us

• Develop parallel algorithms for the library STAPL, a parallel superset of the C++ Standard Template Library• Develop a parallel partitioning framework for arbitrary graphs and unstructured meshes as part of STAPL -- Improved performance of parallel sweep on 3D unstructured meshes by 30% by developing a columnar mesh partitioner• Helped the team to integrate STAPL into a parallel particle transport code leading to a reduction of the mesh construction by a factor of 5• Implement performance and correctness tests for nightly validation of the framework• Interact with a team of 10 people: meetings on the design of the STAPL library, research presentations and mentoring of intern students

College Station, Tx, Us

Developed a parallel uncollided flux algorithm for a particle transport code-- Improved the accuracy of the particle detection

Master Of Science, Computer Science

Master, General Engineering; Computer Science