Xeon Phi Acceleration of Domain Decomposition Iterations via Heterogeneous Active Messages

被引：2

作者：

Maly, Lukas ^{[1
,2
]}

Zapletal, Jan ^{[1
,2
]}

Merta, Michal ^{[1
,2
]}

Cermak, Martin ^{[1
]}

机构：

[1] VSB Tech Univ Ostrava, Natl Supercomp Ctr IT4Innovat, 17 Listopadu 15-2172, Ostrava 70833, Czech Republic

[2] VSB Tech Univ Ostrava, Dept Appl Math, 17 Listopadu 15-2172, Ostrava 70833, Czech Republic

来源：

INTERNATIONAL CONFERENCE OF NUMERICAL ANALYSIS AND APPLIED MATHEMATICS (ICNAAM 2017) | 2018年 / 1978卷

关键词：

D O I：

10.1063/1.5043963

中图分类号：

O29 [应用数学];

学科分类号：

070104 ;

摘要：

We present an acceleration strategy for a domain-decomposition iterative solver based on local Schur complements with respect to the skeleton of the computational domain. In finite element tearing and interconnecting (FETI) this results in an iterative application of a large number of dense matrices. For the offload of such kernels to the Intel (R) Xeon Phi (TM) coprocessors we use the Heterogeneous Active Messages (HAM) library. A simple load balancing strategy is presented to efficiently utilize both the host CPU and the available coprocessors during individual iterations.

引用

页数：4

共 11 条

[1]

Intel Corporation, 2016, SYMM COMM INT SCIF I

[2]

Intel Corporation, 2016, INT C COMP 17 0 DEV

[3]

Intel Corporation, 2016, INTR HET STREAMS LIB

[4]

Intel Corporation, 2016, INT XEON PHI COPR SO

[5]

Jeffers J., 2014, HIGH PERFORMANCE PAR

[6]

Maly L., 2017, P 5 INT C PAR DISTR

[7]

Merta M., 2016, ADV ENG SOFTWARE

[8]

Noack M., 2014, TECH REP, P14

[9]

Noack M., 2017, HETEROGENEOUS ACTIVE

[10] A Unified Programming Model for Intra- and Inter-Node Offloading on Xeon Phi Clusters [J].

Noack, Matthias ;

Wende, Florian ;

Steinke, Thomas ;

Cordes, Frank .

SC14: INTERNATIONAL CONFERENCE FOR HIGH PERFORMANCE COMPUTING, NETWORKING, STORAGE AND ANALYSIS, 2014, :203-214

← 1 2 →