DESC0024749

Problem Statement There is an ongoing trend to move HPC applications to execute in public cloud computing platforms. The combination of containersĺ portability, reproducibility, and potential for high performance makes them an attractive option for running HPC applications in the cloud. However, when transitioning from traditional HPC clusters to a cloud environment, it is crucial to be aware that public cloud communications are susceptible to many security attacks. Since HPC applications are often used to process sensitive data, such as medical, financial, or scientific simulation, the security of these applications is a rising concern. While the use of containers is able to mitigate most of the intra-node security challenges, inter-node communication still poses a significant security concern. Therefore, we must adopt an approach to ensure secure communication in this new environment. While encrypting data seems like the straightforward solution for this, providing cybersecurity support for distributed HPC applications that often span multiple physical compute nodes without performance degradation is challenging. Furthermore, there is currently no available option in container orchestrations to leverage offloading encryption onto emerging Smart-NICs, such as NVIDIA Data Processing Units (DPUs), to ensure support for performance isolation. Proposed Problem Resolution and Phase-I Activities: To address the above challenges, X- ScaleSolutions proposes to design and develop a novel ôSecure-MPIö library based on the MVAPICH MPI library, a derivative of the MPICH library and funded through ASCR PModels Projects for over a decade. As many existing HPC applications do not use containers, we propose a common solution for secure communication with and without containers. The project will involve R&D activities along the following directions: 1) Design and implement a generic mechanism for selecting encryption schemes with minimal performance degradation 2) Optimized designs supporting encryption schemes with point- to-point operations; 3) Optimized designs supporting encryption schemes with collective operations; 4) Optimized designs for offloading encryption to accelerators (GPU/DPU); 5) Supporting security isola- tion in collocated and overlapped partial allocations; 6) Evaluation of Secure-MPI library using various container engines, container orchestrations, and container network interface 7) Systematic testing to ensure proper integration of the proposed designs. The product will be evaluated for a range of HPC workloads for performance and scalability while addressing security needs. point-to-point, and Tasks 1 & 3, part of tasks 6 & 7 will be carried out in Phase-I. Commercial Applications and Other Benefits: Secure-MPI will provide much-needed secure communication for HPC applications running on existing and emerging HPC systems and cloud computing facilities without significant performance degradation. This can lead to a significant boost in performance and scalability of many security or privacy sensitive applications that use MPI libraries running on a wide pool of systems with various cost-effective resource sharing usage models. We will provide a commercial version of this middleware for use by DOE Labs, industries, and cloud providers worldwide.