---

              pEVSL (parallel EigenValue Slicing Library)

---

Latest changes made on:

main data structures for parallel computing

1. evsl_comm This structure contains all process informations in the parallel environment and all the MPI communicators. There are 3 MPI_comm's that are 1) the global communicator (MPI_WORLD) 2), group communicator (for each slice), and 3) group lead communicator, i.e., communicator for all the rank-0 processes in all the groups

2. evsl_parcsr This structure is for parallel CSR matrices. Each process hold a block of rows, say Ai = A(r_i:r_{i+1}-1,:) According to row/column partitioning, Ai is split into a local part (in the diagonal block) and an external part (off block diagonal), which are 2 CSR matrix, namely diag and offd. Currently, the only operation supported is parcsr_Matvec. A structure called commHandle is in evsl_parcsr, which takes care of all communications for matvec. Non-block communications are used in matvec to overlap computaions and communications.

3. evsl_parvec This structure is for parallel vectors.

FILES:

• INC: [all header files]

  • pevsl.h : functino prototypes and constant definitions

  • pevsl_blaslapack.h : BLAS/LAPACK macros

  • pevsl_def.h : Macro definitions

  • pevsl_int.h : internal header with constant definitions and prototypes

  • pevsl_struct.h : pEVSL structs

• SRC:

  • cheblanNr.c : Polynomial Filtered no-restart Lanczos

  • chebpol.c : Computing and applying polynomial filters

  • kpmdos.c : Compute DOS by KPM methods

  • landos.c : Function to use Lanczos method for approximating DOS for the generalized eigenvalue problem

  • lantrbnd.c : A more robust algorithm to give bounds of spectrum based on TR Lanczos

  • lspolapprox.c : Least squares polynomial approximation to a matrix function

  • miscla.c : Miscellaneous la functions

  • parcsr.c : Functinos related to parallel csr structure

  • parcsrmv.c : Parallel csr matrix vector products

  • parvec.c : Parallel vector related functions

  • parvecs.c : Parallel multi-vector related functions

  • pevsl.c : PEVSL interface functions

  • simpson.c : Simpson integrater

  • spmat.c : Space matrix operations

  • spslicer.c : Spectrum slicing

  • stats.c : Used to track various statistics (time taken by various operations)

  • utils.c : Utility functions

  • vector.c : Vector related functions

---

INSTALL: every effort is made so that the only thing to modify is the file makefile.in

  [0. need BLAS/LAPACK and MPI]
  1. modify makefile.in (may need to change MKLRoot, BLAS/LAPACK/MPI linking)
  2. make clean; make

  make in the directory SRC will create the library libpevsl.a

  In directory TESTS/ you will find a makefile to create 3 directorys
  * Gen_Lap              [Generalized Eigenvalue Laplacian]
  * MM                   [Matrix market form]
  * Lap                  [Laplacian]

Run:

      mpirun -np 10 ./test.ex -n 400 ./test.ex -nx 20 -ny 20 -nz 20 -nslices 5  -a 0.6 -b 1.2


      [5 slices, 2 processes for each slice]

= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =
 nx 20 ny 20 nz 20, nslices 5, a = 6.000000e-01 b = 1.200000e+00
= = = = = = = = = = = = = = = = SLICES  FOUND = = = = = = = = = = = = = = = = = = = =
 Slice 0: [6.000000000000000e-01, 7.431715857928962e-01]
 Slice 1: [7.431715857928962e-01, 8.719359679839922e-01]
 Slice 2: [8.719359679839922e-01, 9.904952476238122e-01]
 Slice 3: [9.904952476238122e-01, 1.098549274637319e+00]
 Slice 4: [1.098549274637319e+00, 1.200000000000000e+00]
= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =
 Eigenvalue bounds: (0.000000e+00, 1.200000e+01)
 Eigenvalue count 126.44, estimated count for each slice 31
= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =
 Group  0 (NP=2): [6.00e-01, 7.43e-01], eval computed 18 (18), polydeg  61, its 210
 Group  1 (NP=2): [7.43e-01, 8.72e-01], eval computed 27 (27), polydeg  74, its 270
 Group  2 (NP=2): [8.72e-01, 9.90e-01], eval computed 27 (27), polydeg  86, its 270
 Group  3 (NP=2): [9.90e-01, 1.10e+00], eval computed 19 (19), polydeg  99, its 210
 Group  4 (NP=2): [1.10e+00, 1.20e+00], eval computed 27 (27), polydeg 110, its 270
= = = = = = = = = = = = = = = = = = = SLICE  0 = = = = = = = = = = = = = = = = = = = =
 Timing (sec):
   Create Comms  0.000628,   Create Parcsr 0.002328,   Create Slicer 0.000001
   Apply Slicer  0.245428,   FiltEig Solve 1.105189
 Memory (MB):
   Total 20.94,  Peak 18.60 
= = = = = = = = = = = = = = = = = = = SLICE  1 = = = = = = = = = = = = = = = = = = = =
 Timing (sec):
   Create Comms  0.000217,   Create Parcsr 0.002418,   Create Slicer 0.000001
   Apply Slicer  0.245796,   FiltEig Solve 1.689088
 Memory (MB):
   Total 21.52,  Peak 19.18 
= = = = = = = = = = = = = = = = = = = SLICE  2 = = = = = = = = = = = = = = = = = = = =
 Timing (sec):
   Create Comms  0.000205,   Create Parcsr 0.002302,   Create Slicer 0.000002
   Apply Slicer  0.245934,   FiltEig Solve 1.950552
 Memory (MB):
   Total 21.52,  Peak 19.18 
= = = = = = = = = = = = = = = = = = = SLICE  3 = = = = = = = = = = = = = = = = = = = =
 Timing (sec):
   Create Comms  0.000242,   Create Parcsr 0.002334,   Create Slicer 0.000002
   Apply Slicer  0.245882,   FiltEig Solve 1.717661
 Memory (MB):
   Total 21.01,  Peak 18.67 
= = = = = = = = = = = = = = = = = = = SLICE  4 = = = = = = = = = = = = = = = = = = = =
 Timing (sec):
   Create Comms  0.000216,   Create Parcsr 0.002417,   Create Slicer 0.000002
   Apply Slicer  0.245852,   FiltEig Solve 2.392627
 Memory (MB):
   Total 21.52,  Peak 19.18 
= = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = = =

      ```