The High Performance Conjugate Gradients Benchmark ( HPCG benchmark ) is a supercomputing benchmark test proposed by Michael Heroux from Sandia National Laboratories , and Jack Dongarra and Piotr Luszczek from the University of Tennessee .
4-428: It is intended to model the data access patterns of real-world applications such as sparse matrix calculations, thus testing the effect of limitations of the memory subsystem and internal interconnect of the supercomputer on its computing performance. Because it is internally I/O bound (the data for the benchmark resides in main memory as it is too large for processor caches), HPCG testing generally achieves only
8-527: A database or other repository . Two fundamental types of data access exist: Data access crucially involves authorization to access different data repositories. Data access can help distinguish the abilities of administrators and users. For example, administrators may have the ability to remove, edit and add data, while general users may not even have "read" rights if they lack access to particular information. Historically, each repository (including each different database, file system , etc.), might require
12-613: A tiny fraction of the peak FLOPS the computer could theoretically deliver. HPCG is intended to complement benchmarks such as the LINPACK benchmarks that put relatively little stress on the internal interconnect. The source of the HPCG benchmark is available on GitHub . As of June 2018, the Summit supercomputer held the top spot in the HPCG performance rankings, followed by the Sierra and
16-518: The K computer . In June of 2020, Summit was superseded by Fugaku with a speed of 16.0 HPCG-petaflops (an increase of 540%). Summit is currently 4th, LUMI 3rd and Frontier 2nd. Data access Data access is a generic term referring to a process which has both an IT-specific meaning and other connotations involving access rights in a broader legal and/or political sense. In the former it typically refers to software and activities related to storing, retrieving, or acting on data housed in
#19980