PREMA.CDT3D

From crtc.cs.odu.edu
Revision as of 06:13, 29 March 2018 by Ctsolakis (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search

Decoupled Approach for CDT3D

  1. Take advantage of CDT3D performance and mesh quality capability and create a ‘coarse’ mesh
  2. Decompose the mesh into N subdomains (N > P)
  3. Distribute and refine each subdomain in parallel with no communication.

Domain Decomposition

Due to the absence of appropriate software simple heuristics have been employed

Decoupled partitioned.png Decoupled partitioned2.png

Preliminary Results (May 2017)

Initial Mesh: 880k Tetrahedra; Final Mesh 150M Tetrahedra ; Number of Subdomains 640
empty User Handler threads Total Time Subdomain Migrations
Pthreads (32) - 3288 -
PREMA
64 processes, 128 cores without ILB
1 579 0
PREMA
64 processes, 128 cores with ILB.
1 333 123
PREMA
42 processes, 126 cores with ILB.
2 263 215
PREMA
25 processes, 125 cores with ILB.
3 197 136