Scalapack Performance for LU decomposition

Scalapack blocksize dependency:

• Measurements:

  
  

  

  
  

• good agreement with Scalapack User Guide:
  • generally use 64
  • larger for large SMP machines

Scalapack processor grid dependency:

• Measurements:

  
  

  

  
  

• use 1xn grid
• Scalapack User's Guide:
  • choose $P_r << P_c$, one-dimensional for $P < 8$

Scalapack versus MLIB:

• Times:

  
  

  

  
  

• Efficiency:

  
  

  

  
  

• no Scalapack value for 1 cpu (needs 64bit version)
• no MLIB value for 16 cpus (cluster configuration)
• MLIB better for 1 and 2 cpus,
• Scalapack has better scalability
• Scalapack needs environment variable MPI_2BCOPY=1
  (otherwise performance degradation to 25 %!)