Thank you Mathew. Now, it is all making sense to me. From data file ksp_ex45_N511_gpu_2.txt
KSPSolve (53%) + KSPSetup (0%) = PCSetup (16%) + PCApply (37%). However, you said “So an iteration would mostly consist of MatMult + PCApply, with some vector work” The MalMult event is 4 %. How does this event figure into the above equation; if preconditioning (MatMult + PCApply) is included in KSPSolve? Best, Karthik. From: Matthew Knepley <[email protected]> Date: Wednesday, 29 September 2021 at 10:58 To: "Chockalingam, Karthikeyan (STFC,DL,HC)" <[email protected]> Cc: Barry Smith <[email protected]>, "[email protected]" <[email protected]> Subject: Re: [petsc-users] %T (percent time in this phase) On Wed, Sep 29, 2021 at 5:52 AM Karthikeyan Chockalingam - STFC UKRI <[email protected]<mailto:[email protected]>> wrote: That was helpful. I would like to provide some additional details of my run on cpus and gpus. Please find the following attachments: 1. graph.pdf a plot showing overall time and various petsc events. 2. ksp_ex45_N511_cpu_6.txt data file of the log_summary 3. ksp_ex45_N511_gpu_2.txt data file of the log_summary I used the following petsc options for cpu mpirun -n 6 ./ex45 -log_summary -da_grid_x 511 -da_grid_y 511 -da_grid_z 511 -dm_mat_type mpiaij -dm_vec_type mpi -ksp_type cg -pc_type bjacobi -ksp_monitor and for gpus mpirun -n 1 ./ex45 -log_summary -da_grid_x 511 -da_grid_y 511 -da_grid_z 511 -dm_mat_type mpiaijcusparse -dm_vec_type mpicuda -ksp_type cg -pc_type bjacobi -ksp_monitor to run the following problem https://petsc.org/release/src/ksp/ksp/tutorials/ex45.c.html From the above code, I see is there no individual function called KSPSetUp(), so I gather KSPSetDM, KSPSetComputeInitialGuess, KSPSetComputeRHS, kSPSetComputeOperators all are timed together as KSPSetUp. For this example, is KSPSetUp time and KSPSolve time mutually exclusive? No, KSPSetUp() will be contained in KSPSolve() if it is called automatically. In your response you said that “PCSetUp() time may be in KSPSetUp() or it maybe in PCApply() it depends on how much of the preconditioner construction can take place early, so depends exactly on the preconditioner used.” I don’t see a explicit call to PCSetUp() or PCApply() in ex45; so for this particular preconditioner (bjacobi) how can I tell how they are timed? They are all inside KSPSolve(). If you have a preconditioned linear solve, the oreconditioning happens during the iteration. So an iteration would mostly consist of MatMult + PCApply, with some vector work. I am hoping to time KSP solving and preconditioning mutually exclusively. I am not sure that concept makes sense here. See above. Thanks, Matt Kind regards, Karthik. From: Barry Smith <[email protected]<mailto:[email protected]>> Date: Tuesday, 28 September 2021 at 19:19 To: "Chockalingam, Karthikeyan (STFC,DL,HC)" <[email protected]<mailto:[email protected]>> Cc: "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>> Subject: Re: [petsc-users] %T (percent time in this phase) On Sep 28, 2021, at 12:11 PM, Karthikeyan Chockalingam - STFC UKRI <[email protected]<mailto:[email protected]>> wrote: Thanks for Barry for your response. I was just benchmarking the problem with various preconditioner on cpu and gpu. I understand, it is not possible to get mutually exclusive timing. However, can you tell if KSPSolve time includes both PCSetup and PCApply? And if KSPSolve and KSPSetup are mutually exclusive? Likewise for PCSetUp and PCApply. If you do not call KSPSetUp() separately from KSPSolve() then its time is included with KSPSolve(). PCSetUp() time may be in KSPSetUp() or it maybe in PCApply() it depends on how much of the preconditioner construction can take place early, so depends exactly on the preconditioner used. So yes the answer is not totally satisfying. The one thing I would recommend is to not call KSPSetUp() directly and then KSPSolve() will always include the total time of the solve plus all setup time. PCApply will contain all the time to apply the preconditioner but may also include some setup time. Barry Best, Karthik. From: Barry Smith <[email protected]<mailto:[email protected]>> Date: Tuesday, 28 September 2021 at 16:56 To: "Chockalingam, Karthikeyan (STFC,DL,HC)" <[email protected]<mailto:[email protected]>> Cc: "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>> Subject: Re: [petsc-users] %T (percent time in this phase) On Sep 28, 2021, at 10:55 AM, Karthikeyan Chockalingam - STFC UKRI <[email protected]<mailto:[email protected]>> wrote: Hello, I ran ex45 in the KPS tutorial, which is a 3D finite-difference Poisson problem. I noticed from the output from using the flag -log_summary that for various events their respective %T (percent time in this phase) do not add up to 100 but rather exceeds 100. So, I gather there is some overlap among these events. I am primarily looking at the events KSPSetUp, KSPSolve, PCSetUp and PCSolve. Is it possible to get a mutually exclusive %T or Time for these individual events? I have attached the log_summary output file from my run for your reference. For nested solvers it is tricky to get the times to be mutually exclusive because some parts of the building of the preconditioner is for some preconditioners delayed until the solve has started. It looks like you are using the default preconditioner options which for this example are taking more or less no time since so many iterations are needed. It is best to use -pc_type mg to use geometric multigrid on this problem. Barry Thanks! Karthik. This email and any attachments are intended solely for the use of the named recipients. If you are not the intended recipient you must not use, disclose, copy or distribute this email or any of its attachments and should notify the sender immediately and delete this email from your system. UK Research and Innovation (UKRI) has taken every reasonable precaution to minimise risk of this email or any attachments containing viruses or malware but the recipient should carry out its own virus and malware checks before opening the attachments. UKRI does not accept any liability for any losses or damages which the recipient may sustain due to presence of any viruses. <ksp_ex45_N511_cpu_6.txt> -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener https://www.cse.buffalo.edu/~knepley/<http://www.cse.buffalo.edu/~knepley/>
