I'd like to add that Matlab keeps incorporating the latest sparse direct solvers coming from Tim Davis and his group from TAMU / U. of Florida ( SuiteSparse <http://faculty.cse.tamu.edu/davis/suitesparse.html>) into their new versions. I believe that if the Jacobian is symmetric, current versions of MATLAB will use CHOLMOD, while if it's not, they will use UMFPACK.
This is great because these are solid, state of the art direct solvers; however, as far as I know, there is still no way in Matlab to tune the spparms in order to deactivate their multifrontal / supernodal variants and just use the "simplicial" variants instead. In some testing we did a while ago on the C version of SuiteSparse, the multifrontal and supernodal approaches performed worse on the kind matrices that one typically obtains in power networks. It made sense, because those techniques are essentially trying to find denser blocks in order to use the BLAS, and power systems matrices are just too sparse for that approach to pay off. I hope Matlab implements the KLU solver as an option some day, because my hunch is that KLU is the fastest solver for power systems problems (it was used on Xyce, a SPICE-like simulator). -- Jose L. Marin Gridquant España SL Grupo AIA On Mon, Oct 19, 2015 at 3:04 PM, Ray Zimmerman <r...@cornell.edu> wrote: > I would also mention, for those who are interested, that version 5.1 of > MATPOWER includes a wrapper function mplinsolve() > <http://www.pserc.cornell.edu//matpower/docs/ref/matpower5.1/mplinsolve.html> > that > allows you to choose between different linear solvers for computing the > Newton update step in the MIPS interior-point OPF algorithm. Currently this > includes only Matlab’s built-in \ operator or the optional PARDISO. > > If I remember correctly, for the Newton-Raphson power flow, I stuck with > using Matlab’s \ operator directly rather than mplinsolve() > <http://www.pserc.cornell.edu//matpower/docs/ref/matpower5.1/mplinsolve.html>, > because even for the largest systems I tried, there was little or no > advantage to PARDISO, and the extra overhead was noticeable on small > systems. > > Ray > > > On Oct 19, 2015, at 1:05 AM, Shruti Rao <sra...@asu.edu> wrote: > > Thank you Dr. Abhyankar for the guidance. I appreciate your time and > effort. > > Shruti > > On Sun, Oct 18, 2015 at 10:02 PM, Abhyankar, Shrirang G. <abhy...@anl.gov> > wrote: > >> Shruti, >> MATPOWER does use “\” operator for the linear solves. However note >> that, internally, MATLAB does perform some sort of matrix reordering to >> reduce the fill-ins in the factored matrix. For instance, UMFPACK uses an >> approximate minimum degree reordering scheme by default. >> >> Shri >> >> From: Shruti Rao <sra...@asu.edu> >> Reply-To: MATPOWER discussion forum <matpowe...@list.cornell.edu> >> Date: Sunday, October 18, 2015 at 8:31 PM >> To: MATPOWER discussion forum <matpowe...@list.cornell.edu> >> Subject: Re: Question about sparsity-based implementation in MATPower >> >> Thank you Dr. Abhyakar, >> >> My main aim was to confirm that MATPower uses the inbuilt "\" to solve >> the matrix equations and not Tinney or some other form of reordering and >> then LU factorization followed by forward,backward substitutions. From your >> response I assume that it is true that MATpower uses "\" right? >> >> Thank you for your response. >> >> >> >> On Sun, Oct 18, 2015 at 6:27 PM, Abhyankar, Shrirang G. <abhy...@anl.gov> >> wrote: >> >>> Hi Shruti, >>> The direct linear solver used by MATLAB depends on the symmetry of the >>> Jacobian matrix. For MATPOWER test cases that have symmetric Jacobians (due >>> to inactive taps), a Cholesky factorization is used (LL^T = A). For cases >>> that lead to non-symmetric Jacobian, MATLAB uses UMFPACK for performing the >>> linear solve. >>> >>> Shri >>> >>> From: Shruti Rao <sra...@asu.edu> >>> Reply-To: MATPOWER discussion forum <matpowe...@list.cornell.edu> >>> Date: Sunday, October 18, 2015 at 5:37 PM >>> To: MATPOWER discussion forum <matpowe...@list.cornell.edu> >>> Subject: Question about sparsity-based implementation in MATPower >>> >>> Greetings MATPower community, >>> >>> I had a question about the way sparsity-based techniques are used in the >>> Newton-Raphson solver of the power flow algorithm in MATPower. >>> >>> I ran the code step-by-step and from my understanding, the way the >>> sparsity of the Jacobian matrix is exploited is that it is created as a >>> MATLAB "sparse" matrix wherein only the non-zeros are stored with the >>> respective matrix positions and then the MATLAB operator "\" is invoked >>> while calculating dx = -(J \ F); where J is the Jacobian and F is the >>> vector of mismatches. >>> >>> MATLAB "\" by default exploits the sparsity of the matrix by using a LU >>> solver. The kind of solver "\" uses actually depends on the matrix >>> structure if it is diagonal/tridiagonal/banded and so on (Flowchart >>> obtained from Mathworks website attached in the email). I assume based on >>> the typical structure of the Jacobian that an LU solver is most likely to >>> be chosen. >>> >>> Is my understanding correct or am I missing something out? Thank you for >>> your time and effort. >>> >>> >>> -- >>> Best Regards, >>> Shruti Dwarkanath Rao >>> >>> Graduate Research Assistant >>> School of Electrical, Computer and Energy Engineering >>> Arizona State University >>> Tempe, AZ, 85281 >>> 650 996 0116 >>> >>> >> >> >> -- >> Best Regards, >> Shruti Dwarkanath Rao >> >> Graduate Research Assistant >> School of Electrical, Computer and Energy Engineering >> Arizona State University >> Tempe, AZ, 85281 >> 650 996 0116 >> >> > > > -- > Best Regards, > Shruti Dwarkanath Rao > > Graduate Research Assistant > School of Electrical, Computer and Energy Engineering > Arizona State University > Tempe, AZ, 85281 > 650 996 0116 > > >
