Parallel Implementation and Performance Analysis of ILU Family Preconditioners

Bo Yang


In this work, we implemented the the ILU(0), ILU(k), ILUT  and decoupled block-wise ILU(k) preconditioners for Krylovlinear solvers on GPU. They have Favorable parallel performance with over 5 times speedup obtained in general.  Following ILU-preconditioner Characteristics are verified:

–ILU(0) has better parallel performance than ILUT.

–ILUT has better convergence than ILU(0).

–Better convergence is obtained by higher k or higher block size because of more fill-in entries.

–Favorable parallel performance is obtained for lower k and lower block size.

We can see that ILU-preconditioned linear solver on GPU workstations is a reasonable acceleration solution for reservoir simulation.