[petsc-users] approaches to reduce computing time

Tue Nov 12 15:34:16 CST 2013

On Tue, Nov 12, 2013 at 3:22 PM, Roc Wang <pengxwang at hotmail.com> wrote:

>
>
> ------------------------------
> Date: Tue, 12 Nov 2013 14:59:30 -0600
> Subject: Re: [petsc-users] approaches to reduce computing time
> From: knepley at gmail.com
> To: pengxwang at hotmail.com
> CC: jedbrown at mcs.anl.gov; petsc-users at mcs.anl.gov
>
> On Tue, Nov 12, 2013 at 2:48 PM, Roc Wang <pengxwang at hotmail.com> wrote:
>
>
>
> ------------------------------
> Date: Tue, 12 Nov 2013 14:22:35 -0600
> Subject: Re: [petsc-users] approaches to reduce computing time
> From: knepley at gmail.com
> To: pengxwang at hotmail.com
> CC: jedbrown at mcs.anl.gov; petsc-users at mcs.anl.gov
>
> On Tue, Nov 12, 2013 at 2:14 PM, Roc Wang <pengxwang at hotmail.com> wrote:
>
> Thanks Jed,
>
> I have questions about load balance and PC type below.
>
> > From: jedbrown at mcs.anl.gov
> > To: pengxwang at hotmail.com; petsc-users at mcs.anl.gov
> > Subject: Re: [petsc-users] approaches to reduce computing time
> > Date: Sun, 10 Nov 2013 12:20:18 -0700
> >
> > Roc Wang <pengxwang at hotmail.com> writes:
> >
> > > Hi all,
> > >
> > > I am trying to minimize the computing time to solve a large sparse
> matrix. The matrix dimension is with m=321 n=321 and p=321. I am trying to
> reduce the computing time from two directions: 1 finding a Pre-conditioner
> to reduce the number of iterations which reduces the time numerically, 2
> requesting more cores.
> > >
> > > ----For the first method, I tried several methods:
> > > 1 default KSP and PC,
> > > 2 -ksp_type fgmres -ksp_gmres_restart 30 -pc_type ksp -ksp_pc_type
> jacobi,
> > > 3 -ksp_type lgmres -ksp_gmres_restart 40 -ksp_lgmres_augment 10,
> > > 4 -ksp_type lgmres -ksp_gmres_restart 50 -ksp_lgmres_augment 10,
> > > 5 -ksp_type lgmres -ksp_gmres_restart 40 -ksp_lgmres_augment 10
> -pc_type asm (PCASM)
> > >
> > > The iterations and timing is like the following with 128 cores
> requested:
> > > case# iter timing (s)
> > > 1 1436 816
> > > 2 3 12658
> > > 3 1069 669.64
> > > 4 872 768.12
> > > 5 927 513.14
> > >
> > > It can be seen that change -ksp_gmres_restart and -ksp_lgmres_augment
> can help to reduce the iterations but not the timing (comparing case 3 and
> 4). Second, the PCASM helps a lot. Although the second option is able to
> reduce iterations, the timing increases very much. Is it because more
> operations are needed in the PC?
> > >
> > > My questions here are: 1. Which direction should I take to select
> > > -ksp_gmres_restart and -ksp_lgmres_augment? For example, if larger
> > > restart with large augment is better or larger restart with smaller
> > > augment is better?
> >
> > Look at the -log_summary. By increasing the restart, the work in
> > KSPGMRESOrthog will increase linearly, but the number of iterations
> > might decrease enough to compensate. There is no general rule here
> > since it depends on the relative expense of operations for your problem
> > on your machine.
> >
> > > ----For the second method, I tried with -ksp_type lgmres
> -ksp_gmres_restart 40 -ksp_lgmres_augment 10 -pc_type asm with different
> number of cores. I found the speedup ratio increases slowly when more than
> 32 to 64 cores are requested. I searched the milling list archives and
> found that I am very likely running into the memory bandwidth bottleneck.
> http://www.mail-archive.com/petsc-users@mcs.anl.gov/msg19152.html:
> > >
> > > # of cores iter timing
> > > 1 923 19541.83
> > > 4 929 5897.06
> > > 8 932 4854.72
> > > 16 924 1494.33
> > > 32 924 1480.88
> > > 64 928 686.89
> > > 128 927 627.33
> > > 256 926 552.93
> >
> > The bandwidth issue has more to do with using multiple cores within a
> > node rather than between nodes. Likely the above is a load balancing
> > problem or bad communication.
>
> I use DM to manage the distributed data.  The DM was created by calling
> DMDACreate3d() and let PETSc decide the local number of nodes in each
> direction. To my understand the load of each core is determined at this
> stage.   If the load balance is done when DMDACreate3d() is called and use
> PETSC_DECIDE option? Or how should make the load balanced after DM is
> created?
>
>
> We do not have a way to do fine-grained load balancing for the DMDA since
> it is intended for very simple topologies. You can see
> if it is load imbalance from the division by running with a cube that is
> evenly divisible with a cube number of processes.
>
>    Matt
>
> So, I have nothing to do to make the load balanced if I use DMDA?  Would
> you please take a look at the attached log summary files and give me some
> suggestions on how to improve the speedup ratio? Thanks.
>
>
> Please try what I suggested above. And it looks like there is a little
> load imbalance
>
> Roc----So if the domain is a cube, then the number of the processors is
> better to be like 2^3=8, 3^3=9, 4^4 =16, and so on, right?
>

I want you to try this to eliminate load imbalance as a reason for poor
speedup. I don't think it is, but we will see.

> I am also wondering whether the physical boundary type effects the load
> balance? Since freed node, Dirichlet node and Neumann node has different
> number of neighbors?
>
> VecAXPY              234 1.0 1.0124e+00 3.4 1.26e+08 1.1 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0 15290
>
> VecAXPY              234 1.0 4.2862e-01 3.6 6.37e+07 1.1 0.0e+00 0.0e+00 0.0e+00  0  0  0  0  0   0  0  0  0  0 36115
>
>
> although it is not limiting the speedup. The time imbalance is really
> strange. I am guessing other jobs are running on this machine.
>
>
> Roc----The code was run a cluster. There should be other jobs were
> running. Do you mean those jobs affect the load balance of my job or speed
> of the cluster?  I am just trying to improve the scalability of the code,
> but really don't know what's the reason that the speedup ratio decreases
> so quickly? Thanks.
>

Yes, other people running can definitely screw up speedup and cause
imbalance. Usually timing runs are made with dedicated time.

Your VecAXPY and MatMult are speeding up just fine. It is reductions which
are killing your computation.
You should switch to a more effective preconditioner, so you can avoid all
those dot products. Also, you
might try something like BiCG with fewer dot products.

   Matt

>    Matt
>
>
> >
> > > My question here is: Is there any other PC can help on both reducing
> iterations and increasing scalability? Thanks.
> >
> > Always send -log_summary with questions like this, but algebraic
> multigrid is a good place to start.
>
> Please take a look at the attached log file, they are for 128 cores and
> 256 cores, respectively.  Based on the log files, what should be done to
> increase the scalability? Thanks.
>
>
>
>
> --
> 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
>
>
>
>
> --
> 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
>

-- 
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20131112/1b52c9b2/attachment-0001.html>