Impact of the implementation of MPI point-to-point communications on the performance of two general sparse solvers

Abstract

The current concept for high-level radioactive waste disposal at Yucca Mountain is for the waste to be placed in underground tunnels (or drifts) in the middle of a thick unsaturated zone.Flow modeling and field testing have shown that not all flow encountering a drift will seep into the drift. The underlying reason for the diversion of unsaturated flow around a drift is that capillary forces in the fractures and matrix prevent water entry into the drift unless the capillary pressure in the rock decreases sufficiently to allow for gravity forces to overcome the capillary barrier. As a result of the capillary barrier effect, flow tends to be diverted around the drift, affecting the flow pattern beneath the drift. For some distance beneath the drift, water saturation and flux are reduced. This drift shadow zone is much more pronounced in the fractures than in the matrix due to dominance of gravity over capillary forces in the fractures. Moving downward, away from the drift, the shadow zone asymptotically re-equilibrates to the undisturbed flow conditions due to capillary forces. The behavior of radionuclide transport in this zone of reduced flow is investigated here because this will affect the amount of time required * Corresponding author. Fax: +1-510-486-6115; e-mail: JEHouseworth@lbl.gov 2 for radionuclides to penetrate the unsaturated zone. The delay of radionuclide movement in the geosphere is one aspect of the potential repository system that could limit public exposure to radioactive waste.The behavior of flow and transport are calculated using a two-dimensional, drift-scale dual permeability model extending to nine drift diameters below the potential waste emplacement drift. The flow model is first compared with an analytical model for a single-continuum. Then the dual-continuum flow model is investigated with respect to drift-scale and mountain-scale property sets. Transport calculations are performed for a wide range of flow conditions and for different aqueous radionuclides and colloids. Findings indicate that transport times for dissolved or colloidal material released from a drift without seepage are several orders of magnitude longer than if the releases occurred in the undisturbed flow field. Furthermore, the calculations indicate that the transport rate for radionuclides released in the drift shadow is relatively insensitive to flow rates in the fractures, but is sensitive to the flow rate in the matrix.Key Words: Unsaturated zone flow, fractured rock, radionuclide transport, Yucca Mountain, numerical modeling. Introduction Hydrogeology at Yucca MountainYucca Mountain consists of alternating layers of welded and nonwelded volcanic tuffs. The potential repository horizon lies in the unsaturated zone, approximately 300 m below the ground surface and 300 m above the water table. This horizon consists of welded tuffs composed of a 3 dense porous rock matrix, which is highly fractured. Fracture permeability in the welded tuffs is on the order of darcies (~10 -12 m 2 ), and the rock matrix permea...

Parallel sparse LU factorization on different message passing platforms

“…For instance, Amestoy et al studied the impact of the MPI buffering implementation on the performance of sparse matrix solvers [36]. Hunold et al proposed multilevel hierarchical matrix multiplication to improve the application performance on the PC cluster [37].…”

A survey of direct methods for sparse linear systems

“…Both methods have the same total communication volume but the multifrontal method requires fewer messages. In a subsequent paper (Amestoy, Duff, L'Excellent and Li 2003a), they show how MPI implementations affect both solvers. Li (2005) provides an overview of all three SuperLU variants: (1) the left-looking sequential SuperLU, (2) the left-looking parallel shared-memory SuperLU MT, and (3) the right-looking parallel distributed-memory Su-perLU DIST.…”

“…Amestoy et al (2001b) compare their two distributed-memory approaches and software: SuperLU DIST, which relies on a synchronous supernodal method with static pivoting and iterative refinement (discussed in Section 9.2), and MUMPS, which uses an asynchronous multifrontal method with partial threshold and delayed pivoting. Amestoy et al (2003a) report their experience of using MPI point-topoint communications on the performance of their respective solvers. They present challenges and solutions on the use of buffered asynchronous message transmission and reception in MPI.…”

Parallel sparse LU factorization on second-class message passing platforms

“…For instance, Amestoy et al studied the impact of the MPI buffering implementation on the performance of sparse matrix solvers [2]. Chakrabarti and Yelick investigated application-controlled consistency mechanisms to minimize synchronization and communication overhead for solving the Gröbner basis problem on distributed memory machines [4].…”