< BK21 플러스 BEST 정보기술 사업단 세미나 개최 안내 >

개최일시 : 2018 년 12월 21일 (금) 13:00 ~ 15:00

개최장소 : 제4공학관 D402호

세미나 제목 : IO intensive computing and graph processing

내용 :

With the GPUs gaining its popularity, the importance of advanced IO is arising along. From

POWER8 processor, advanced IO interfaces called CAPI and NVLINK have been introduced to provide a rich amount of bandwidth to accelerators. This allows to build a much faster system with GPUs, or FPGAs as high-performance accelerators.

This talk will cover IBM’s recent effort on IO intensive computing with CAPI (Coherent Accelerator Processor Interface) along with some work on graph processing acceleration using CAPI technology. As an example, I will introduce ExtraV, an out-of-memory graph processing accelerator using CAPI. In ExtraV, a CAPI-enabled FPGA accelerator sits between a processor and a storage device. Using a high-level programming model called graph virtualization, it hides the storage format of the graph data and performs the traversals for the processor. Also, it applies a few optimizations under the hidden layer such as bandwidth amplification and graph multi-versioning.

If time permits, I’d like to additionally talk about a recent work on a novel algorithm for compression of graph data. By using a specially tailored variable-length coding for graph structure and a random access index, the algorithm provides a much higher compression ratio for graphs, with constant-time random access.

강연자 성함&직함 / 소속 : Dr. Jinho Lee, IBM Research, Austin

초청자 : 전기전자공학과 교수 노원우