Den Jan 8, 2014 kl. 1:00 PM skrev Martin Sandve Alnæs: > You can call mesh.init explicitly in your code to measure them separately? >
Yes, just did. I replaced the DirichletBC lines in the script by mesh.init() and re-ran. For me a full mesh.init() leads to memory consumption of 1053900 KB, or 1 GB! Compared to the Laplacian matrix of 8 MB. Is this reproducible? M > Martin > > > On 8 January 2014 12:43, Mikael Mortensen <[email protected]> wrote: > Hi, > > I am currently comparing a fenics based solver to some other open source > Navier-Stokes solvers and I just realized that I was using much more memory, > perhaps as much as a factor 3. I always thought it was the matrices, vectors > and Krylov solvers that were consuming memory, but on inspection I see that > memory is totally dominated by the boundary conditions (mesh connectivity?). > This was quite surprising to me, so I guess my question is - is this a well > known issue? > > I have made a small test for memory use of a Poisson solver to illustrate. > Running it on one cpu returns the table showing the increment in RSS memory > use and total memory use for each new function down the script. > > In [1]: run memorytest.py > RSS memory Increment Total > Starting weight of dolfin 75988 KB 75988 KB > Mesh 16760 KB 92748 KB > FunctionSpace 41600 KB 134348 KB > Matrix assembly 7792 KB 142140 KB > Function 304 KB 142444 KB > Solve 868 KB 143312 KB > Boundary condition 321428 KB 464740 KB > New solve 0 KB 464740 KB > > Creating and applying the Dirichlet boundary condition calls mesh.init(D) and > mesh.init(D-1, D). I think this is were the extra memory comes in. The memory > cost is nearly 5 times more than for all the other stuff! Any comments? > > Mikael > > > > from dolfin import * > from commands import getoutput > from os import getpid > set_log_active(False) > > def getRSSMemoryUsage(): > mypid = getpid() > return eval(getoutput("ps -o rss %s" % mypid).split()[1]) > > class MemoryUsage: > def __init__(self, s): > self.memory = 0 > info_blue(' '*21+'RSS memory Increment Total') > self(s) > > def __call__(self, s): > self.prev = self.memory > self.memory = getRSSMemoryUsage() > #self.memory = memory_usage(False)[0] > info_blue('{0:26s} {1:10d} KB {2:10d} KB'.format(s, > self.memory-self.prev, self.memory)) > > memoryusage = MemoryUsage('Starting weight of dolfin') > > mesh = UnitCubeMesh(40, 40, 40) > > memoryusage("Mesh") > > V = FunctionSpace(mesh, 'CG', 1) > > memoryusage('FunctionSpace') > > u, v = TrialFunction(V), TestFunction(V) > A = assemble(inner(grad(u), grad(v))*dx) > > memoryusage("Matrix assembly") > > b = assemble(Constant(6)*v*dx) > u = Function(V) > > memoryusage("Function") > > solve(A, u.vector(), b, 'gmres', 'hypre_amg') > > memoryusage('Solve') > > bc = DirichletBC(V, Constant(0), DomainBoundary()) > bc.apply(A, b) > > memoryusage('Boundary condition') > > solve(A, u.vector(), b, 'gmres', 'hypre_amg') > > memoryusage('New solve') > > _______________________________________________ > fenics mailing list > [email protected] > http://fenicsproject.org/mailman/listinfo/fenics >
_______________________________________________ fenics mailing list [email protected] http://fenicsproject.org/mailman/listinfo/fenics
