Chen Wang is Computing’s sixth Sidney Fernbach Postdoctoral Fellow in the Computing Sciences. Named for a former LLNL Director of Computation, this competitive fellowship is awarded to exceptional postdocs who demonstrate the potential for significant achievements in computational mathematics, computer science, data science, or scientific computing. Fellows work in the Computing Directorate on their own research agenda and with a mentor’s guidance.
Chen joined the Lab in September after completing his PhD in Computer Science at the University of Illinois Urbana-Champaign. Originally from Qingdao on the east coast of China, he earned bachelor’s and master’s degrees in Computer Science from Hainan University and Tianjin University, respectively, before moving to the U.S. in 2017. Chen interned at LLNL in 2019 and since then has maintained a collaboration with Center for Applied Scientific Computing (CASC) computer scientist Kathryn Mohror, who is mentoring him during his fellowship. “We share the same research interest in improving the I/O performance of HPC systems and have conducted work on parallel I/O tracing and analysis,” he states.
The collaborators developed a tool called Recorder, whose algorithm compresses I/O traces of large-scale runs and can therefore trace I/O activities on extremely large scales. They also came up with a method that verifies an application’s I/O consistency requirements while running on a file system. This investigation revealed something crucial: “We showed that most HPC applications do not require a strong consistency model such as POSIX to run correctly,” Chen explains. “This assumption is made by many I/O researchers but has never been confirmed experimentally before.” (POSIX, which stands for Portable Operating System Interface, is a set of standards and definitions for ensuring operating systems’ compatibility.)
Chen plans to spend his fellowship upending the POSIX-centric I/O programming paradigm. “Instead of enforcing a single standard consistency model, we advocate for a standard synchronization interface and return the choice of consistency models to storage system designers,” he says, adding that lessons learned from memory consistency models are often wasted when designing storage consistency models. Chen intends to connect these models—to compare their similarities and differences and ultimately propose a unified framework for specifying storage consistency models. “I want to answer questions about the best consistency model for HPC applications, and how much performance improvement we can gain,” states Chen.
Chen first learned about HPC in a parallel computing class a decade ago, and he immediately recognized the I/O subsystem’s importance and the inherent challenges of parallel storage. He points out, “As parallel storage researchers, on the one hand, we need to work with HPC applications to understand their I/O behaviors to better serve their I/O needs. On the other hand, we need to tackle I/O research issues such as the best design choice of a parallel storage system.”
As a Fernbach Fellow, Chen looks forward to the opportunity to pursue his ideas alongside CASC researchers. He says, “My research requires access to LLNL’s powerful supercomputers, and in turn, the outcome of my research should be beneficial for the design of the I/O subsystem of those machines.” Mohror adds, “I am so happy that Chen chose to come to LLNL to do his research. The approach he is taking is a game changer. He’s rethinking the fundamentals of how we do I/O, and his work is poised to have huge impacts on code performance.”