From: KIRK, BENJAMIN (JSCEG) (NASA) <benjamin.kirk1@na...>  20051014 13:27:21

The alive time is the total elapsed time between creating and deleting the performance log. when the alive time and active time are disparate it is an indication that the processor is spending some idle time waiting, or there is work going on that is not specifically logged. One thought is that ~100k degrees of freedom is rather small for 10 processors. You might try a larger mesh and see if that changes anything. The communication latency is essentially independent of problem size, and it can become the dominant factor when passing small messages. What happens is that for small problems the computations are actually so fast that the latency required to send and receive a message becomes the dominant factor in performance. Another idea is use PETSc's internal logging capabilities to glean more insight into what is going on. Simply run your application with the log_summary option and PETSc will spit out a performance log at the end of the run. Ben Original Message From: libmeshusersadmin@... [mailto:libmeshusersadmin@...] On Behalf Of Min Xu Sent: Thursday, October 13, 2005 9:05 PM To: libmeshusers@... Subject: [Libmeshusers] Performance of libmesh and PETSC in a linux cluster Hi, In an attempt to test the performance of libmesh and petsc in a linux cluster, a 2D timedependent diffusion equation is solved on a uniform mesh of size 320x320 and the solution is performed on 74 time steps. In the first case, 5 processors are used. The diffusion system is: EquationSystems n_systems()=1 System "Diffusion" Type "TransientLinearImplicit" Variables="u" Finite Element Types="LAGRANGE" Approximation Orders="FIRST" n_dofs()=103041 n_local_dofs()=20850 n_constrained_dofs()=0 n_vectors()=3 The assembling takes 4.8251s for one time step and the total 74 petsc solutions takes 396.8s. In the second case, total 10 processors are used. The diffusion system is: EquationSystems n_systems()=1 System "Diffusion" Type "TransientLinearImplicit" Variables="u" Finite Element Types="LAGRANGE" Approximation Orders="FIRST" n_dofs()=103041 n_local_dofs()=10378 n_constrained_dofs()=0 n_vectors()=3 The assembling takes 2.38s for one time step and the total 74 petsc solutions takes 208.2s. All the above times are active times. This proves that both libmesh and petsc scales well with the number of processors. But one thing puzzles me most is the alive time reported is 1226.03s which is 6 times of the petsc solution time. What does the extra time spent come from? Anyone has some idea? Best, Min  This SF.Net email is sponsored by: Power Architecture Resource Center: Free content, downloads, discussions, and more. http://solutions.newsforge.com/ibmarch.tmpl _______________________________________________ Libmeshusers mailing list Libmeshusers@... https://lists.sourceforge.net/lists/listinfo/libmeshusers 