Re: Remeshing and code speedup
Understood. I think you will find that a 2000x2000 2D problem (assuming you have enough RAM) will not take anything like 7.5 h. > On Jul 24, 2018, at 3:49 PM, Carsten Langrock wrote: > > I tried a few solvers for this 1D diffusion problem and it does indeed appear > that pySparse is faster than the rest. Sadly, faster doesn’t equate to fast … > about 4 seconds for a 2000 grid, 440 time steps nonlinear diffusion process > using a single sweep. For the second part, which opens the left boundary and > forces it to zero, it takes another 9.5 seconds when choosing 3 sweeps. And > that’s on a fast computer. > > Thanks again for pointing out the use of different solvers. > > Carsten > > _ > Dipl.-Phys. Carsten Langrock, Ph.D. > > Senior Research Scientist > Edward L. Ginzton Laboratory, Rm. 202 > Stanford University > > 348 Via Pueblo Mall > 94305 Stanford, CA > > Tel. (650) 723-0464 > Fax (650) 723-2666 > > Ginzton Lab Shipping Address: > James and Anna Marie Spilker Engineering and Applied Sciences Building > 04-040 > 348 Via Pueblo Mall > 94305 Stanford, CA > _ > >> On Jul 24, 2018, at 6:11 AM, Guyer, Jonathan E. Dr. (Fed) >> wrote: >> >> FiPy still does not support remeshing. >> >> As Dario said, choice of solver can make a big difference. I've not used >> PyAMG much, but PySparse is dramatically faster than SciPy. PyTrilinos is >> slower than PySparse, but enables you to solve in parallel. >> >> I've also found that 2D problems solve much better than the 1D performance >> would lead you to believe. There's just a lot of overhead in setting up the >> problem and the Python communication with the lower-level libraries. >> >> On Jul 23, 2018, at 6:44 PM, Carsten Langrock wrote: >> >> Hi, >> >> Thanks for the help with getting FiPy running under Linux! I am trying to >> re-create a 1D nonlinear diffusion problem for which we have C++ code that >> uses the implicit Thomas algorithm based on >> >> J. Weickert, B. Romerny, M. Viergever, "Efficient and Reliable Schemes >> for Nonlinear Diffusion Filtering”, IEEE transactions on Image Processing, >> vol.7, N03, page 398, March 1998 >> >> I have been able to get results in FiPy that match this code very closely >> which was a great start. Our C++ code uses a fixed number of spatial points >> and a fixed time step, but re-meshes space to most efficiently use the size >> of the array; it increases the spatial step size by 2 whenever the >> concentration at a particular point reaches a set threshold. I tried >> implementing this in FiPy as well, but haven’t had much luck so far. I saw >> an old mailing-list entry from 2011 where a user was told that FiPy wasn’t >> meant to do remeshing. Is that still the case? >> >> I’d imagine one would somehow need to update the Grid1D object with the new >> ‘dx’, but since the CellVariable that holds the solution was initialized >> with that mesh object, I am not sure that such a change would propagate in a >> sensible fashion. I think I know how to map the value of the CellVariable to >> account for the change in ‘dx’ by >> >> array_size = 2000 >> phi.value = numpy.concatenate((phi.value[1:array_size/2:2], >> numpy.zeros(1500))) >> >> for the case when the initial variable holds 2000 spatial points. Maybe >> there’s a more elegant way, but I think this works in principle. >> >> Another question would be execution speed. Right now, even when not plotting >> the intermediate solutions, it takes many seconds on a very powerful >> computer to run a simple diffusion problem. I am probably doing something >> really wrong. I wasn’t expecting the code to perform as well as the C++ >> code, but I had hoped to come within an order of magnitude. Are there ways >> to optimize the performance? Maybe select a particularly clever solver? If >> someone could point me into the right direction that’d be great. In the end, >> I would like to expand the code into 2D, but given the poor 1D performance, >> I don’t think that this would be feasible at this point. >> >> Thanks, >> Carsten >> >> _ >> Dipl.-Phys. Carsten Langrock, Ph.D. >> >> Senior Research Scientist >> Edward L. Ginzton Laboratory, Rm. 202 >> Stanford University >> >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> >> Tel. (650) 723-0464 >> Fax (650) 723-2666 >> >> Ginzton Lab Shipping Address: >> James and Anna Marie Spilker Engineering and Applied Sciences Building >> 04-040 >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> _ >> ___ >> fipy mailing list >> fipy@nist.gov >> http://www.ctcms.nist.gov/fipy >> [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] >> >> >> >> ___ >> fipy mailing list >> fipy@nist.gov >> http://www.ctcms.nist.gov/fipy >> [ NIST internal ONLY: https://em
Re: Remeshing and code speedup
Thanks for the links. I had come across the first one, but I haven’t found any major speedup using the —inline option on this 1D problem. Initially I thought that remeshing would be fairly simple, not unlike what our C++ code does. It wouldn’t speed things up necessarily since our code uses a constant number of points, but it would increase the accuracy since you are reducing the computational window to only contain areas with non-zero concentration (in the case of a diffusion problem). >From the response by the maintainers of FiPy it would seem that this isn’t as >straightforward as I thought. I’ll give it another go, though, and would >report back if it works. The overhead might be pretty large, let’s see. Carsten _Dipl.-Phys. Carsten Langrock, Ph.D. Senior Research Scientist Edward L. Ginzton Laboratory, Rm. 202 Stanford University 348 Via Pueblo Mall 94305 Stanford, CA Tel. (650) 723-0464 Fax (650) 723-2666 Ginzton Lab Shipping Address: James and Anna Marie Spilker Engineering and Applied Sciences Building 04-040 348 Via Pueblo Mall 94305 Stanford, CA _ > On Jul 24, 2018, at 9:19 AM, Drew Davidson wrote: > > > Hi Carsten, > > Have you looked at: > > > https://www.ctcms.nist.gov/fipy/documentation/EFFICIENCY.html > > > > > http://nbviewer.jupyter.org/github/wd15/fipy-efficiency/blob/master/notebooks/FiPy-IPython.ipynb > > > > > https://www.mail-archive.com/fipy@nist.gov/msg03180.html > thread > > > > I struggled to install weave and get --inline working. I didn't use it much > yet, but my initial impression was the speedup was modest. > > > Remeshing seems like a great possibility for making problems such as > examples.phase.anisotropy run faster (use higher mesh density at the > interface), but I am afraid that FiPy architecture (lazy evaluation etc) > could interfere. I hope you share what you find with remeshing. > > > Thanks > > > On Tue, Jul 24, 2018 at 10:56 AM Carsten Langrock > wrote: > >> >> Thanks for pointing out the performance order of the solvers. I’ll try to >> get pySparse to work to compare it other solvers. It’s also good to know >> that I shouldn’t give up on 2D just yet;-) >> >> >> Regards, >> Carsten >> >> >> _Dipl.-Phys. Carsten Langrock, Ph.D. >> >> Senior Research Scientist >> Edward L. Ginzton Laboratory, Rm. 202 >> Stanford University >> >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> >> Tel. (650) 723-0464 >> Fax (650) 723-2666 >> >> Ginzton Lab Shipping Address: >> James and Anna Marie Spilker Engineering and Applied Sciences Building >> 04-040 >> >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> _ >> >> >> >> >> >>> On Jul 24, 2018, at 6:11 AM, Guyer, Jonathan E. Dr. (Fed) >>> >>> wrote: >>> >>> >>> FiPy still does not support remeshing. >>> >>> >>> As Dario said, choice of solver can make a big difference. I've not used >>> PyAMG much, but PySparse is dramatically faster than SciPy. PyTrilinos is >>> slower than PySparse, but enables you to solve in parallel. >>> >>> >>> I've also found that 2D problems solve much better than the 1D performance >>> would lead you to believe. There's just a lot of overhead in setting up the >>> problem and the Python communication with the lower-level libraries. On Jul 23, 2018, at 6:44 PM, Carsten Langrock wrote: Hi, Thanks for the help with getting FiPy running under Linux! I am trying to re-create a 1D nonlinear diffusion problem for which we have C++ code that uses the implicit Thomas algorithm based on J. Weickert, B. Romerny, M. Viergever, "Efficient and Reliable Schemes for Nonlinear Diffusion Filtering”, IEEE transactions on Image Processing, vol.7, N03, page 398, March 1998 I have been able to get results in FiPy that match this code very closely which was a great start. Our C++ code uses a fixed number of spatial points and a fixed time step, but re-meshes space to most efficiently use the size of the array; it increases the spatial step size by 2 whenever the concentration at a particular point reaches a set threshold. I tried implementing this in FiPy as well, but haven’t had much luck so far. I saw an old mailing-list entry from 2011 where a user was told that FiPy wasn’t meant to do remeshing. Is that still the case? I’d imagine one would somehow need to update the Grid1D object with the new ‘dx’, but since the CellVariable that holds the solution was initialized with that mesh object, I am not sure that such a change would propagate in a sensible fashion. I think I know how to map the value of the CellVariable to account for the change in ‘dx’ by array_size = 2000 phi.value = numpy.conc
Re: Remeshing and code speedup
I tried a few solvers for this 1D diffusion problem and it does indeed appear that pySparse is faster than the rest. Sadly, faster doesn’t equate to fast … about 4 seconds for a 2000 grid, 440 time steps nonlinear diffusion process using a single sweep. For the second part, which opens the left boundary and forces it to zero, it takes another 9.5 seconds when choosing 3 sweeps. And that’s on a fast computer. Thanks again for pointing out the use of different solvers. Carsten _Dipl.-Phys. Carsten Langrock, Ph.D. Senior Research Scientist Edward L. Ginzton Laboratory, Rm. 202 Stanford University 348 Via Pueblo Mall 94305 Stanford, CA Tel. (650) 723-0464 Fax (650) 723-2666 Ginzton Lab Shipping Address: James and Anna Marie Spilker Engineering and Applied Sciences Building 04-040 348 Via Pueblo Mall 94305 Stanford, CA _ > On Jul 24, 2018, at 6:11 AM, Guyer, Jonathan E. Dr. (Fed) > wrote: > > > FiPy still does not support remeshing. > > > As Dario said, choice of solver can make a big difference. I've not used > PyAMG much, but PySparse is dramatically faster than SciPy. PyTrilinos is > slower than PySparse, but enables you to solve in parallel. > > > I've also found that 2D problems solve much better than the 1D performance > would lead you to believe. There's just a lot of overhead in setting up the > problem and the Python communication with the lower-level libraries. >> >> >> On Jul 23, 2018, at 6:44 PM, Carsten Langrock wrote: >> >> >> Hi, >> >> >> Thanks for the help with getting FiPy running under Linux! I am trying to >> re-create a 1D nonlinear diffusion problem for which we have C++ code that >> uses the implicit Thomas algorithm based on >> >> >> J. Weickert, B. Romerny, M. Viergever, "Efficient and Reliable Schemes >> for Nonlinear Diffusion Filtering”, IEEE transactions on Image Processing, >> vol.7, N03, page 398, March 1998 >> >> >> I have been able to get results in FiPy that match this code very closely >> which was a great start. Our C++ code uses a fixed number of spatial points >> and a fixed time step, but re-meshes space to most efficiently use the size >> of the array; it increases the spatial step size by 2 whenever the >> concentration at a particular point reaches a set threshold. I tried >> implementing this in FiPy as well, but haven’t had much luck so far. I saw >> an old mailing-list entry from 2011 where a user was told that FiPy wasn’t >> meant to do remeshing. Is that still the case? >> >> >> I’d imagine one would somehow need to update the Grid1D object with the new >> ‘dx’, but since the CellVariable that holds the solution was initialized >> with that mesh object, I am not sure that such a change would propagate in a >> sensible fashion. I think I know how to map the value of the CellVariable to >> account for the change in ‘dx’ by >> >> >> array_size = 2000 >> phi.value = numpy.concatenate((phi.value[1:array_size/2:2], >> numpy.zeros(1500))) >> >> >> for the case when the initial variable holds 2000 spatial points. Maybe >> there’s a more elegant way, but I think this works in principle. >> >> >> Another question would be execution speed. Right now, even when not plotting >> the intermediate solutions, it takes many seconds on a very powerful >> computer to run a simple diffusion problem. I am probably doing something >> really wrong. I wasn’t expecting the code to perform as well as the C++ >> code, but I had hoped to come within an order of magnitude. Are there ways >> to optimize the performance? Maybe select a particularly clever solver? If >> someone could point me into the right direction that’d be great. In the end, >> I would like to expand the code into 2D, but given the poor 1D performance, >> I don’t think that this would be feasible at this point. >> >> >> Thanks, >> Carsten >> >> >> _ >> Dipl.-Phys. Carsten Langrock, Ph.D. >> >> >> Senior Research Scientist >> Edward L. Ginzton Laboratory, Rm. 202 >> Stanford University >> >> >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> >> >> Tel. (650) 723-0464 >> Fax (650) 723-2666 >> >> >> Ginzton Lab Shipping Address: >> James and Anna Marie Spilker Engineering and Applied Sciences Building >> 04-040 >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> _ >> ___ >> fipy mailing list >> fipy@nist.gov >> http://www.ctcms.nist.gov/fipy >> [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] >> > > > > > > > > ___ > fipy mailing list > fipy@nist.gov > http://www.ctcms.nist.gov/fipy > [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] > ___ fipy mailing list fipy@nist.gov http://www.ctcms.nist.gov/fipy [ NIST internal ONLY: ht
Re: Remeshing and code speedup
Hi Carsten, Have you looked at: https://www.ctcms.nist.gov/fipy/documentation/EFFICIENCY.html http://nbviewer.jupyter.org/github/wd15/fipy-efficiency/blob/master/notebooks/FiPy-IPython.ipynb https://www.mail-archive.com/fipy@nist.gov/msg03180.html thread I struggled to install weave and get --inline working. I didn't use it much yet, but my initial impression was the speedup was modest. Remeshing seems like a great possibility for making problems such as examples.phase.anisotropy run faster (use higher mesh density at the interface), but I am afraid that FiPy architecture (lazy evaluation etc) could interfere. I hope you share what you find with remeshing. Thanks On Tue, Jul 24, 2018 at 10:56 AM Carsten Langrock wrote: > Thanks for pointing out the performance order of the solvers. I’ll try to > get pySparse to work to compare it other solvers. It’s also good to know > that I shouldn’t give up on 2D just yet;-) > > Regards, > Carsten > > _ > *Dipl.-Phys. Carsten Langrock, Ph.D.* > > Senior Research Scientist > Edward L. Ginzton Laboratory, Rm. 202 > Stanford University > > 348 Via Pueblo Mall > 94305 Stanford, CA > > Tel. (650) 723-0464 > Fax (650) 723-2666 > > Ginzton Lab Shipping Address: > James and Anna Marie Spilker Engineering and Applied Sciences Building > 04-040 > 348 Via Pueblo Mall > 94305 Stanford, CA > _ > > On Jul 24, 2018, at 6:11 AM, Guyer, Jonathan E. Dr. (Fed) < > jonathan.gu...@nist.gov> wrote: > > FiPy still does not support remeshing. > > As Dario said, choice of solver can make a big difference. I've not used > PyAMG much, but PySparse is dramatically faster than SciPy. PyTrilinos is > slower than PySparse, but enables you to solve in parallel. > > I've also found that 2D problems solve much better than the 1D performance > would lead you to believe. There's just a lot of overhead in setting up the > problem and the Python communication with the lower-level libraries. > > > On Jul 23, 2018, at 6:44 PM, Carsten Langrock > wrote: > > Hi, > > Thanks for the help with getting FiPy running under Linux! I am trying to > re-create a 1D nonlinear diffusion problem for which we have C++ code that > uses the implicit Thomas algorithm based on > > J. Weickert, B. Romerny, M. Viergever, "Efficient and Reliable Schemes > for Nonlinear Diffusion Filtering”, IEEE transactions on Image Processing, > vol.7, N03, page 398, March 1998 > > I have been able to get results in FiPy that match this code very closely > which was a great start. Our C++ code uses a fixed number of spatial points > and a fixed time step, but re-meshes space to most efficiently use the size > of the array; it increases the spatial step size by 2 whenever the > concentration at a particular point reaches a set threshold. I tried > implementing this in FiPy as well, but haven’t had much luck so far. I saw > an old mailing-list entry from 2011 where a user was told that FiPy wasn’t > meant to do remeshing. Is that still the case? > > I’d imagine one would somehow need to update the Grid1D object with the > new ‘dx’, but since the CellVariable that holds the solution was > initialized with that mesh object, I am not sure that such a change would > propagate in a sensible fashion. I think I know how to map the value of the > CellVariable to account for the change in ‘dx’ by > > array_size = 2000 > phi.value = numpy.concatenate((phi.value[1:array_size/2:2], > numpy.zeros(1500))) > > for the case when the initial variable holds 2000 spatial points. Maybe > there’s a more elegant way, but I think this works in principle. > > Another question would be execution speed. Right now, even when not > plotting the intermediate solutions, it takes many seconds on a very > powerful computer to run a simple diffusion problem. I am probably doing > something really wrong. I wasn’t expecting the code to perform as well as > the C++ code, but I had hoped to come within an order of magnitude. Are > there ways to optimize the performance? Maybe select a particularly clever > solver? If someone could point me into the right direction that’d be great. > In the end, I would like to expand the code into 2D, but given the poor 1D > performance, I don’t think that this would be feasible at this point. > > Thanks, > Carsten > > _ > Dipl.-Phys. Carsten Langrock, Ph.D. > > Senior Research Scientist > Edward L. Ginzton Laboratory, Rm. 202 > Stanford University > > 348 Via Pueblo Mall > 94305 Stanford, CA > > Tel. (650) 723-0464 > Fax (650) 723-2666 > > Ginzton Lab Shipping Address: > James and Anna Marie Spilker Engineering and Applied Sciences Building > 04-040 > 348 Via Pueblo Mall > 94305 Stanford, CA > _ > ___ > fipy mailing list > fipy@nist.gov > http://www.ctcms.nist.gov/fipy > [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] > > > > > ___
Re: Remeshing and code speedup
Thanks for pointing out the performance order of the solvers. I’ll try to get pySparse to work to compare it other solvers. It’s also good to know that I shouldn’t give up on 2D just yet;-) Regards, Carsten _Dipl.-Phys. Carsten Langrock, Ph.D. Senior Research Scientist Edward L. Ginzton Laboratory, Rm. 202 Stanford University 348 Via Pueblo Mall 94305 Stanford, CA Tel. (650) 723-0464 Fax (650) 723-2666 Ginzton Lab Shipping Address: James and Anna Marie Spilker Engineering and Applied Sciences Building 04-040 348 Via Pueblo Mall 94305 Stanford, CA _ > On Jul 24, 2018, at 6:11 AM, Guyer, Jonathan E. Dr. (Fed) > wrote: > > > FiPy still does not support remeshing. > > > As Dario said, choice of solver can make a big difference. I've not used > PyAMG much, but PySparse is dramatically faster than SciPy. PyTrilinos is > slower than PySparse, but enables you to solve in parallel. > > > I've also found that 2D problems solve much better than the 1D performance > would lead you to believe. There's just a lot of overhead in setting up the > problem and the Python communication with the lower-level libraries. >> >> >> On Jul 23, 2018, at 6:44 PM, Carsten Langrock wrote: >> >> >> Hi, >> >> >> Thanks for the help with getting FiPy running under Linux! I am trying to >> re-create a 1D nonlinear diffusion problem for which we have C++ code that >> uses the implicit Thomas algorithm based on >> >> >> J. Weickert, B. Romerny, M. Viergever, "Efficient and Reliable Schemes >> for Nonlinear Diffusion Filtering”, IEEE transactions on Image Processing, >> vol.7, N03, page 398, March 1998 >> >> >> I have been able to get results in FiPy that match this code very closely >> which was a great start. Our C++ code uses a fixed number of spatial points >> and a fixed time step, but re-meshes space to most efficiently use the size >> of the array; it increases the spatial step size by 2 whenever the >> concentration at a particular point reaches a set threshold. I tried >> implementing this in FiPy as well, but haven’t had much luck so far. I saw >> an old mailing-list entry from 2011 where a user was told that FiPy wasn’t >> meant to do remeshing. Is that still the case? >> >> >> I’d imagine one would somehow need to update the Grid1D object with the new >> ‘dx’, but since the CellVariable that holds the solution was initialized >> with that mesh object, I am not sure that such a change would propagate in a >> sensible fashion. I think I know how to map the value of the CellVariable to >> account for the change in ‘dx’ by >> >> >> array_size = 2000 >> phi.value = numpy.concatenate((phi.value[1:array_size/2:2], >> numpy.zeros(1500))) >> >> >> for the case when the initial variable holds 2000 spatial points. Maybe >> there’s a more elegant way, but I think this works in principle. >> >> >> Another question would be execution speed. Right now, even when not plotting >> the intermediate solutions, it takes many seconds on a very powerful >> computer to run a simple diffusion problem. I am probably doing something >> really wrong. I wasn’t expecting the code to perform as well as the C++ >> code, but I had hoped to come within an order of magnitude. Are there ways >> to optimize the performance? Maybe select a particularly clever solver? If >> someone could point me into the right direction that’d be great. In the end, >> I would like to expand the code into 2D, but given the poor 1D performance, >> I don’t think that this would be feasible at this point. >> >> >> Thanks, >> Carsten >> >> >> _ >> Dipl.-Phys. Carsten Langrock, Ph.D. >> >> >> Senior Research Scientist >> Edward L. Ginzton Laboratory, Rm. 202 >> Stanford University >> >> >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> >> >> Tel. (650) 723-0464 >> Fax (650) 723-2666 >> >> >> Ginzton Lab Shipping Address: >> James and Anna Marie Spilker Engineering and Applied Sciences Building >> 04-040 >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> _ >> ___ >> fipy mailing list >> fipy@nist.gov >> http://www.ctcms.nist.gov/fipy >> [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] >> > > > > > > > > ___ > fipy mailing list > fipy@nist.gov > http://www.ctcms.nist.gov/fipy > [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] > ___ fipy mailing list fipy@nist.gov http://www.ctcms.nist.gov/fipy [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ]
Re: Remeshing and code speedup
Dario, Thanks. I was able to install the PyAmg module and use it by adding —pyamg on the command line. The DefaultSolver on my system appears to be PyTrilinos. Sadly, with the —pyamg flag, it took significantly longer for the script to run … and I haven’t yet been able to get pySparse to work; or make the —inline flag go, for that matter. I’ll check on another machine later. Thanks for the help, Carsten _Dipl.-Phys. Carsten Langrock, Ph.D. Senior Research Scientist Edward L. Ginzton Laboratory, Rm. 202 Stanford University 348 Via Pueblo Mall 94305 Stanford, CA Tel. (650) 723-0464 Fax (650) 723-2666 Ginzton Lab Shipping Address: James and Anna Marie Spilker Engineering and Applied Sciences Building 04-040 348 Via Pueblo Mall 94305 Stanford, CA _ > On Jul 24, 2018, at 2:32 AM, Dario Panada wrote: > > > Hi Carsten, > > I'll start by saying I'm not a fipy expert, but have been playing around it > for a few months as part of my PhD project. > > > Regarding your second question. > > > Performance can be improved by switching from the default (SciPy) solver to > one of the others. I use the PyAmg solver that can solve a 100x100x100 3D > mesh with multiple sources and sinks in about 1 minute. I am not running a > particularly powerful computer, just my laptop with 8GB of RAM (about half of > which taken up by the OS, Browser and IntelliJ), so you could probably get > even better performance. > > > Hope this is at least somewhat helpful. :) > > > Kind Regards, > Dario > > > On Tue, Jul 24, 2018 at 12:44 AM Carsten Langrock > wrote: > >> >> Hi, >> >> >> Thanks for the help with getting FiPy running under Linux! I am trying to >> re-create a 1D nonlinear diffusion problem for which we have C++ code that >> uses the implicit Thomas algorithm based on >> >> >> J. Weickert, B. Romerny, M. Viergever, "Efficient and Reliable Schemes >> for Nonlinear Diffusion Filtering”, IEEE transactions on Image Processing, >> vol.7, N03, page 398, March 1998 >> >> >> >> I have been able to get results in FiPy that match this code very closely >> which was a great start. Our C++ code uses a fixed number of spatial points >> and a fixed time step, but re-meshes space to most efficiently use the size >> of the array; it increases the spatial step size by 2 whenever the >> concentration at a particular point reaches a set threshold. I tried >> implementing this in FiPy as well, but haven’t had much luck so far. I saw >> an old mailing-list entry from 2011 where a user was told that FiPy wasn’t >> meant to do remeshing. Is that still the case? >> >> >> I’d imagine one would somehow need to update the Grid1D object with the new >> ‘dx’, but since the CellVariable that holds the solution was initialized >> with that mesh object, I am not sure that such a change would propagate in a >> sensible fashion. I think I know how to map the value of the CellVariable to >> account for the change in ‘dx’ by >> >> >> array_size = 2000 >> phi.value = numpy.concatenate((phi.value[1:array_size/2:2], >> numpy.zeros(1500))) >> >> >> for the case when the initial variable holds 2000 spatial points. Maybe >> there’s a more elegant way, but I think this works in principle. >> >> >> Another question would be execution speed. Right now, even when not plotting >> the intermediate solutions, it takes many seconds on a very powerful >> computer to run a simple diffusion problem. I am probably doing something >> really wrong. I wasn’t expecting the code to perform as well as the C++ >> code, but I had hoped to come within an order of magnitude. Are there ways >> to optimize the performance? Maybe select a particularly clever solver? If >> someone could point me into the right direction that’d be great. In the end, >> I would like to expand the code into 2D, but given the poor 1D performance, >> I don’t think that this would be feasible at this point. >> >> >> Thanks, >> Carsten >> >> >> _Dipl.-Phys. Carsten Langrock, Ph.D. >> >> Senior Research Scientist >> Edward L. Ginzton Laboratory, Rm. 202 >> Stanford University >> >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> >> Tel. (650) 723-0464 >> Fax (650) 723-2666 >> >> Ginzton Lab Shipping Address: >> James and Anna Marie Spilker Engineering and Applied Sciences Building >> 04-040 >> >> 348 Via Pueblo Mall >> 94305 Stanford, CA >> _ >> >> >> >> >> ___ >> fipy mailing list >> fipy@nist.gov >> >> http://www.ctcms.nist.gov/fipy >> >> [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy >> ] >> > ___ > fipy mailing list > fipy@nist.gov > http://www.ctcms.nist.gov/fipy > [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] >
Re: Remeshing and code speedup
FiPy still does not support remeshing. As Dario said, choice of solver can make a big difference. I've not used PyAMG much, but PySparse is dramatically faster than SciPy. PyTrilinos is slower than PySparse, but enables you to solve in parallel. I've also found that 2D problems solve much better than the 1D performance would lead you to believe. There's just a lot of overhead in setting up the problem and the Python communication with the lower-level libraries. > On Jul 23, 2018, at 6:44 PM, Carsten Langrock wrote: > > Hi, > > Thanks for the help with getting FiPy running under Linux! I am trying to > re-create a 1D nonlinear diffusion problem for which we have C++ code that > uses the implicit Thomas algorithm based on > > J. Weickert, B. Romerny, M. Viergever, "Efficient and Reliable Schemes > for Nonlinear Diffusion Filtering”, IEEE transactions on Image Processing, > vol.7, N03, page 398, March 1998 > > I have been able to get results in FiPy that match this code very closely > which was a great start. Our C++ code uses a fixed number of spatial points > and a fixed time step, but re-meshes space to most efficiently use the size > of the array; it increases the spatial step size by 2 whenever the > concentration at a particular point reaches a set threshold. I tried > implementing this in FiPy as well, but haven’t had much luck so far. I saw an > old mailing-list entry from 2011 where a user was told that FiPy wasn’t meant > to do remeshing. Is that still the case? > > I’d imagine one would somehow need to update the Grid1D object with the new > ‘dx’, but since the CellVariable that holds the solution was initialized with > that mesh object, I am not sure that such a change would propagate in a > sensible fashion. I think I know how to map the value of the CellVariable to > account for the change in ‘dx’ by > > array_size = 2000 > phi.value = numpy.concatenate((phi.value[1:array_size/2:2], > numpy.zeros(1500))) > > for the case when the initial variable holds 2000 spatial points. Maybe > there’s a more elegant way, but I think this works in principle. > > Another question would be execution speed. Right now, even when not plotting > the intermediate solutions, it takes many seconds on a very powerful computer > to run a simple diffusion problem. I am probably doing something really > wrong. I wasn’t expecting the code to perform as well as the C++ code, but I > had hoped to come within an order of magnitude. Are there ways to optimize > the performance? Maybe select a particularly clever solver? If someone could > point me into the right direction that’d be great. In the end, I would like > to expand the code into 2D, but given the poor 1D performance, I don’t think > that this would be feasible at this point. > > Thanks, > Carsten > > _ > Dipl.-Phys. Carsten Langrock, Ph.D. > > Senior Research Scientist > Edward L. Ginzton Laboratory, Rm. 202 > Stanford University > > 348 Via Pueblo Mall > 94305 Stanford, CA > > Tel. (650) 723-0464 > Fax (650) 723-2666 > > Ginzton Lab Shipping Address: > James and Anna Marie Spilker Engineering and Applied Sciences Building > 04-040 > 348 Via Pueblo Mall > 94305 Stanford, CA > _ > ___ > fipy mailing list > fipy@nist.gov > http://www.ctcms.nist.gov/fipy > [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] ___ fipy mailing list fipy@nist.gov http://www.ctcms.nist.gov/fipy [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ]
Re: Remeshing and code speedup
Hi Carsten, I'll start by saying I'm not a fipy expert, but have been playing around it for a few months as part of my PhD project. Regarding your second question. Performance can be improved by switching from the default (SciPy) solver to one of the others. I use the PyAmg solver that can solve a 100x100x100 3D mesh with multiple sources and sinks in about 1 minute. I am not running a particularly powerful computer, just my laptop with 8GB of RAM (about half of which taken up by the OS, Browser and IntelliJ), so you could probably get even better performance. Hope this is at least somewhat helpful. :) Kind Regards, Dario On Tue, Jul 24, 2018 at 12:44 AM Carsten Langrock wrote: > Hi, > > Thanks for the help with getting FiPy running under Linux! I am trying to > re-create a 1D nonlinear diffusion problem for which we have C++ code that > uses the implicit Thomas algorithm based on > > J. Weickert, B. Romerny, M. Viergever, "Efficient and Reliable Schemes > for Nonlinear Diffusion Filtering”, IEEE transactions on Image Processing, > vol.7, N03, page 398, March 1998 > > I have been able to get results in FiPy that match this code very closely > which was a great start. Our C++ code uses a fixed number of spatial points > and a fixed time step, but re-meshes space to most efficiently use the size > of the array; it increases the spatial step size by 2 whenever the > concentration at a particular point reaches a set threshold. I tried > implementing this in FiPy as well, but haven’t had much luck so far. I saw > an old mailing-list entry from 2011 where a user was told that FiPy wasn’t > meant to do remeshing. Is that still the case? > > I’d imagine one would somehow need to update the Grid1D object with the > new ‘dx’, but since the CellVariable that holds the solution was > initialized with that mesh object, I am not sure that such a change would > propagate in a sensible fashion. I think I know how to map the value of the > CellVariable to account for the change in ‘dx’ by > > array_size = 2000 > phi.value = numpy.concatenate((phi.value[1:array_size/2:2], > numpy.zeros(1500))) > > for the case when the initial variable holds 2000 spatial points. Maybe > there’s a more elegant way, but I think this works in principle. > > Another question would be execution speed. Right now, even when not > plotting the intermediate solutions, it takes many seconds on a very > powerful computer to run a simple diffusion problem. I am probably doing > something really wrong. I wasn’t expecting the code to perform as well as > the C++ code, but I had hoped to come within an order of magnitude. Are > there ways to optimize the performance? Maybe select a particularly clever > solver? If someone could point me into the right direction that’d be great. > In the end, I would like to expand the code into 2D, but given the poor 1D > performance, I don’t think that this would be feasible at this point. > > Thanks, > Carsten > > _ > *Dipl.-Phys. Carsten Langrock, Ph.D.* > > Senior Research Scientist > Edward L. Ginzton Laboratory, Rm. 202 > Stanford University > > 348 Via Pueblo Mall > 94305 Stanford, CA > > Tel. (650) 723-0464 > Fax (650) 723-2666 > > Ginzton Lab Shipping Address: > James and Anna Marie Spilker Engineering and Applied Sciences Building > 04-040 > 348 Via Pueblo Mall > 94305 Stanford, CA > _ > ___ > fipy mailing list > fipy@nist.gov > http://www.ctcms.nist.gov/fipy > [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ] > ___ fipy mailing list fipy@nist.gov http://www.ctcms.nist.gov/fipy [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ]
Remeshing and code speedup
Hi, Thanks for the help with getting FiPy running under Linux! I am trying to re-create a 1D nonlinear diffusion problem for which we have C++ code that uses the implicit Thomas algorithm based on J. Weickert, B. Romerny, M. Viergever, "Efficient and Reliable Schemes for Nonlinear Diffusion Filtering”, IEEE transactions on Image Processing, vol.7, N03, page 398, March 1998 I have been able to get results in FiPy that match this code very closely which was a great start. Our C++ code uses a fixed number of spatial points and a fixed time step, but re-meshes space to most efficiently use the size of the array; it increases the spatial step size by 2 whenever the concentration at a particular point reaches a set threshold. I tried implementing this in FiPy as well, but haven’t had much luck so far. I saw an old mailing-list entry from 2011 where a user was told that FiPy wasn’t meant to do remeshing. Is that still the case? I’d imagine one would somehow need to update the Grid1D object with the new ‘dx’, but since the CellVariable that holds the solution was initialized with that mesh object, I am not sure that such a change would propagate in a sensible fashion. I think I know how to map the value of the CellVariable to account for the change in ‘dx’ by array_size = 2000 phi.value = numpy.concatenate((phi.value[1:array_size/2:2], numpy.zeros(1500))) for the case when the initial variable holds 2000 spatial points. Maybe there’s a more elegant way, but I think this works in principle. Another question would be execution speed. Right now, even when not plotting the intermediate solutions, it takes many seconds on a very powerful computer to run a simple diffusion problem. I am probably doing something really wrong. I wasn’t expecting the code to perform as well as the C++ code, but I had hoped to come within an order of magnitude. Are there ways to optimize the performance? Maybe select a particularly clever solver? If someone could point me into the right direction that’d be great. In the end, I would like to expand the code into 2D, but given the poor 1D performance, I don’t think that this would be feasible at this point. Thanks, Carsten _Dipl.-Phys. Carsten Langrock, Ph.D. Senior Research Scientist Edward L. Ginzton Laboratory, Rm. 202 Stanford University 348 Via Pueblo Mall 94305 Stanford, CA Tel. (650) 723-0464 Fax (650) 723-2666 Ginzton Lab Shipping Address: James and Anna Marie Spilker Engineering and Applied Sciences Building 04-040 348 Via Pueblo Mall 94305 Stanford, CA _ ___ fipy mailing list fipy@nist.gov http://www.ctcms.nist.gov/fipy [ NIST internal ONLY: https://email.nist.gov/mailman/listinfo/fipy ]