Pgas Communication Runtime For Extreme Large Data Computation

SC16: The International Conference for High Performance Computing, Networking, Storage and Analysis Salt Lake City Utah November, 2016(2016)

引用 4|浏览1
暂无评分
摘要
For partitioned global address space (PGAS) run-times, supporting out-of-core data computation is an important issue. Some researchers showed that flash SSDs are useful for out-of-core data computation.In this paper, we introduce ComEx-PM, a PGAS communication runtime. ComEx-PM supports out-of-core data computation using a flash SSD. ComEx-PM launched multiple processes in each node. Memory region of each process is swapped into a flash SSD if total size of the data is out-of-core.We adopted ComEx-PM into Global Arrays (GA), a PGAS runtime library. We executed two programs developed with GA. One program is a dense matrix-matrix multiplication (dgemm) program. The other program is a simple 2D stencil computation program. The dgemm program showed ComEx-PM could run the program with the same performance as the original GA. The stencil program illustrated the overhead of ComEx-PM and how to reduce it.
更多
查看译文
关键词
Out-of-core computing,Flash memories,Parallel programming,PGAS,Runtime library
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要