Lightning Talk: Data Science and AI by Big Memory Supercomputer
TimeSunday, 14 November 202110:30am - 11am CST
DescriptionOne of the biggest issues for high performance data science and AI is the performance and capacity of storage, both for main memory and I/O storage. For the main memory, traditional DRAM solution limits the capacity even though the bandwidth and latency are almost sufficient. About I/O storage, SSD technology provides high performance for local storage, however we need a distributed one for large scale parallel processing.
One of the novel technologies to apply for these issues is the persistent memory (PMEM). PMEM is attachable to CPU directly through ordinary memory bus while it provides much larger capacity than DRAM although the latency is sightly lower. There are several usage of this new device such as large capacity directly accessible memory, coupling with DRAM as a sort of cache to compensate the speed of PMEM, or a local I/O storage device replacing SSD.
In the Center for Computational Sciences (CCS) at University of Tsukuba, we are planning to introduce a new supercomputer named "Cygnus-BD" toward a new method of big data science and AI supported by PMEM technology. We call this machine as Big Memory Supercomputer, where BD of the system name stands for Big Data. We combine every technology supported by PMEM both for large capacity local memory coupled with DRAM and also implementing high performance distributed shared storage as an ad hoc file system. In this talk, I will introduce how PMEM can be applied for such applications and systems as well as introduction of our Cygnus-BD system plan.