ICSA Faculty Talk - 2020 TITLE System Software Challenges for New Hardware and Applications ABSTRACT Emerging hardware, such as non-volatile main memory (NVMM) and accelerator devices, changes the way system software should be designed and implemented, because they are not just an enhanced version of existing devices, but provide new qualitative features and software interfaces. In this talk we first discuss implications of NVMMs in networked systems, because most of the modern computer systems are connected to networks over the Internet, datacenter fabric, or both. We demonstrate co-designing a software stack for both networking and NVMM storage can improve throughput and latency in various workloads and applications. We then discuss implications of SoC-based accelerator devices equipped with general-purpose CPU cores. We demonstrate that bioinformatics applications suffer from slow storage access in the existing software architecture, and a specialized storage stack can significantly improve their performance. Finally, we discuss implications of new applications based on our ongoing work towards practical systems support for graph neural network applications. Feb 20 2020 14.00 - 15.00 ICSA Faculty Talk - 2020 Michio Honda G.03, IF
ICSA Faculty Talk - 2020 TITLE System Software Challenges for New Hardware and Applications ABSTRACT Emerging hardware, such as non-volatile main memory (NVMM) and accelerator devices, changes the way system software should be designed and implemented, because they are not just an enhanced version of existing devices, but provide new qualitative features and software interfaces. In this talk we first discuss implications of NVMMs in networked systems, because most of the modern computer systems are connected to networks over the Internet, datacenter fabric, or both. We demonstrate co-designing a software stack for both networking and NVMM storage can improve throughput and latency in various workloads and applications. We then discuss implications of SoC-based accelerator devices equipped with general-purpose CPU cores. We demonstrate that bioinformatics applications suffer from slow storage access in the existing software architecture, and a specialized storage stack can significantly improve their performance. Finally, we discuss implications of new applications based on our ongoing work towards practical systems support for graph neural network applications. Feb 20 2020 14.00 - 15.00 ICSA Faculty Talk - 2020 Michio Honda G.03, IF
