Since 1978
Published in Sarov (Arzamas-16), Nizhegorodskaya oblast

RUSSIAN FEDERAL
NUCLEAR CENTER -
ALL-RUSSIAN RESEARCH INSTITUTE
OF EXPERIMENTAL PHYSICS
 
 Русский |  English
ABOUT EDITORIAL BOARD PUBLICATION ETHICS RULES FOR AUTHORS AUTHORS ARCHIVE MOST RECENT ISSUE IN NEXT ISSUE PAPER OF THE YEAR




COMMUNICATION SPACE DECOMPOSITION METHOD IN THE IMPLEMENTATION OF MPI SHARED OPERATIONS

S. I. Sapronov
VANT. Ser.: Mat. Mod. Fiz. Proc 2013. Вып.2. С. 78-85.

More often the modern high-performance cluster systems have non- uniform infrastructure when the heterogeneous multi-core computational nodes are connected by various communication features. Such non-uniformity can have a substantial impact on parallel application performance with MPI collective communications: the shared memory exchange at separate nodes should efficiently fit the fast and slow channels exchange providing connection between different nodes. The implementation of new algorithms in shared operations accounting for various communication features is one of the possibilities of message exchange adaptation to new application conditions. This approach is often labor-intensive.
In this article we present an alternative approach providing for the possibility of more efficient application of shared applications algorithms. The principle of method is partition of the communication space in which the shared operation is initially activated (all the processes under communicator control) in homogenous sub-spaces (subcommunicators) with further execution of these operations by parts in separate sub-spaces. Thus, a large set of additional shared algorithms is developed.
First, the advantages of these algorithms lie in taking the merits of the existing algorithms in a homogeneous runtime environment; second, in a concurrent execution of separate parts of shared operations in isolated communication sub-spaces; and third, in a reduction of slow communication number.

Keywords: MPI, parallel computations, algorithms of shared operations.








[ Back ]


 
 
 
© FSUE "RFNC-VNIIEF", 2000-2024