Dear Kwant Develop I have compared the performance of default, mumps solvers in windows and Ubuntu machines. Their performance is very similar without significant gain when using mumps. Do I miss anything in the installation? For example, not installing the mumps. Similar performance is found for the sparse solver. The system I considered has 10^5 to 6X10^5 sites and machine has a quad core i7 cpu with 16G RAM. My installation procedures are those found in the website for pre complled packages. KS Disclaimer: This email (including any attachments) is for the use of the intended recipient only and may contain confidential information and/or copyright material. If you are not the intended recipient, please notify the sender immediately and delete this email and all copies from your system. Any unauthorized use, disclosure, reproduction, copying, distribution, or other form of unauthorized dissemination of the contents is expressly prohibited.
Dear KS,
I have compared the performance of default, mumps solvers in windows and Ubuntu machines. Their performance is very similar without significant gain when using mumps. Do I miss anything in the installation? For example, not installing the mumps. Similar performance is found for the sparse solver. The system I considered has 10^5 to 6X10^5 sites and machine has a quad core i7 cpu with 16G RAM. My installation procedures are those found in the website for pre complled packages.
This is contrary to our experience. Can you please provide more details: * How did you install Kwant in both cases? * What specific script did you run? * What were the execution times that you observed? Thanks, Anton
KS
Disclaimer: This email (including any attachments) is for the use of the intended recipient only and may contain confidential information and/or copyright material. If you are not the intended recipient, please notify the sender immediately and delete this email and all copies from your system. Any unauthorized use, disclosure, reproduction, copying, distribution, or other form of unauthorized dissemination of the contents is expressly prohibited.
Dear Anton, To confirm my observation, I remove the non-essential part of the script to test the mumps solver and default solver. The system is a zigzag graphene ribbon, with 200X200 unit cells, there are 4 atoms in one unit cell so there are 1.6x10^5 atoms. The OS is window 10, kwant is 1.2.2 and python is 3.4.4. the time for calculating 20 smatrix is around 30 minutes for both default solver and mumps. Please find attached the script and the runtime. The first runtime is for mumps. The second runtime is for default. The kwant was installed with the package prepared by Gohlke. For the sparse solver the time is longer, 40 minutes but there is no umfpack installed in the window system. So I do not show the runtime as it may not be accurate. KS On 16/5/2017, 5:51 PM, "anton.akhmerov@gmail.com on behalf of Anton Akhmerov" <anton.akhmerov@gmail.com on behalf of anton.akhmerov+kd@gmail.com> wrote: Dear KS, > I have compared the performance of default, mumps solvers in windows and > Ubuntu machines. Their performance is very similar without significant gain > when using mumps. Do I miss anything in the installation? For example, not > installing the mumps. Similar performance is found for the sparse solver. > The system I considered has 10^5 to 6X10^5 sites and machine has a quad core > i7 cpu with 16G RAM. My installation procedures are those found in the > website for pre complled packages. This is contrary to our experience. Can you please provide more details: * How did you install Kwant in both cases? * What specific script did you run? * What were the execution times that you observed? Thanks, Anton > KS > > > > Disclaimer: This email (including any attachments) is for the use of the > intended recipient only and may contain confidential information and/or > copyright material. If you are not the intended recipient, please notify the > sender immediately and delete this email and all copies from your system. > Any unauthorized use, disclosure, reproduction, copying, distribution, or > other form of unauthorized dissemination of the contents is expressly > prohibited. Disclaimer: This email (including any attachments) is for the use of the intended recipient only and may contain confidential information and/or copyright material. If you are not the intended recipient, please notify the sender immediately and delete this email and all copies from your system. Any unauthorized use, disclosure, reproduction, copying, distribution, or other form of unauthorized dissemination of the contents is expressly prohibited.
Dear KS, On a linux workstation the ratio of runtimes is around 2.3 in favor of MUMPS, so the problem seems to pertain to your installation or to the windows installations in general. Can anyone using Kwant on windows try the attached script? (I reduced the number of points so that it doesn't take 1/2 hour and removed some parts not crucial to the benchmark) Best, Anton On Tue, May 16, 2017 at 7:18 PM, Prof. CHAN Kwok Sum <apkschan@cityu.edu.hk> wrote:
Dear Anton, To confirm my observation, I remove the non-essential part of the script to test the mumps solver and default solver. The system is a zigzag graphene ribbon, with 200X200 unit cells, there are 4 atoms in one unit cell so there are 1.6x10^5 atoms. The OS is window 10, kwant is 1.2.2 and python is 3.4.4. the time for calculating 20 smatrix is around 30 minutes for both default solver and mumps. Please find attached the script and the runtime. The first runtime is for mumps. The second runtime is for default. The kwant was installed with the package prepared by Gohlke. For the sparse solver the time is longer, 40 minutes but there is no umfpack installed in the window system. So I do not show the runtime as it may not be accurate. KS
On 16/5/2017, 5:51 PM, "anton.akhmerov@gmail.com on behalf of Anton Akhmerov" <anton.akhmerov@gmail.com on behalf of anton.akhmerov+kd@gmail.com> wrote:
Dear KS,
> I have compared the performance of default, mumps solvers in windows and > Ubuntu machines. Their performance is very similar without significant gain > when using mumps. Do I miss anything in the installation? For example, not > installing the mumps. Similar performance is found for the sparse solver. > The system I considered has 10^5 to 6X10^5 sites and machine has a quad core > i7 cpu with 16G RAM. My installation procedures are those found in the > website for pre complled packages.
This is contrary to our experience. Can you please provide more details: * How did you install Kwant in both cases? * What specific script did you run? * What were the execution times that you observed?
Thanks, Anton
> KS > > > > Disclaimer: This email (including any attachments) is for the use of the > intended recipient only and may contain confidential information and/or > copyright material. If you are not the intended recipient, please notify the > sender immediately and delete this email and all copies from your system. > Any unauthorized use, disclosure, reproduction, copying, distribution, or > other form of unauthorized dissemination of the contents is expressly > prohibited.
Disclaimer: This email (including any attachments) is for the use of the intended recipient only and may contain confidential information and/or copyright material. If you are not the intended recipient, please notify the sender immediately and delete this email and all copies from your system. Any unauthorized use, disclosure, reproduction, copying, distribution, or other form of unauthorized dissemination of the contents is expressly prohibited.
From your email, I think it is indeed kwant.smatrix and kmumps.smatrix
Dear KS, I think there might be a misunderstanding: kwant has right only now two solvers, one using mumps and one using scipy.sparse. If mumps is present, the default solver *points to mumps*, so in this case kwant.smatrix is identical to kmumps.smatrix in your code! that you compare, is this right? (you mention the scipy solver was very slow). Best, Michael On 18-05-17 09:56, Anton Akhmerov wrote:
Dear KS,
On a linux workstation the ratio of runtimes is around 2.3 in favor of MUMPS, so the problem seems to pertain to your installation or to the windows installations in general. Can anyone using Kwant on windows try the attached script? (I reduced the number of points so that it doesn't take 1/2 hour and removed some parts not crucial to the benchmark)
Best, Anton
On Tue, May 16, 2017 at 7:18 PM, Prof. CHAN Kwok Sum <apkschan@cityu.edu.hk> wrote:
Dear Anton, To confirm my observation, I remove the non-essential part of the script to test the mumps solver and default solver. The system is a zigzag graphene ribbon, with 200X200 unit cells, there are 4 atoms in one unit cell so there are 1.6x10^5 atoms. The OS is window 10, kwant is 1.2.2 and python is 3.4.4. the time for calculating 20 smatrix is around 30 minutes for both default solver and mumps. Please find attached the script and the runtime. The first runtime is for mumps. The second runtime is for default. The kwant was installed with the package prepared by Gohlke. For the sparse solver the time is longer, 40 minutes but there is no umfpack installed in the window system. So I do not show the runtime as it may not be accurate. KS
On 16/5/2017, 5:51 PM, "anton.akhmerov@gmail.com on behalf of Anton Akhmerov" <anton.akhmerov@gmail.com on behalf of anton.akhmerov+kd@gmail.com> wrote:
Dear KS,
> I have compared the performance of default, mumps solvers in windows and > Ubuntu machines. Their performance is very similar without significant gain > when using mumps. Do I miss anything in the installation? For example, not > installing the mumps. Similar performance is found for the sparse solver. > The system I considered has 10^5 to 6X10^5 sites and machine has a quad core > i7 cpu with 16G RAM. My installation procedures are those found in the > website for pre complled packages.
This is contrary to our experience. Can you please provide more details: * How did you install Kwant in both cases? * What specific script did you run? * What were the execution times that you observed?
Thanks, Anton
> KS > > > > Disclaimer: This email (including any attachments) is for the use of the > intended recipient only and may contain confidential information and/or > copyright material. If you are not the intended recipient, please notify the > sender immediately and delete this email and all copies from your system. > Any unauthorized use, disclosure, reproduction, copying, distribution, or > other form of unauthorized dissemination of the contents is expressly > prohibited.
Disclaimer: This email (including any attachments) is for the use of the intended recipient only and may contain confidential information and/or copyright material. If you are not the intended recipient, please notify the sender immediately and delete this email and all copies from your system. Any unauthorized use, disclosure, reproduction, copying, distribution, or other form of unauthorized dissemination of the contents is expressly prohibited.
Dear Anton, Thank you for your reply. I note you use kwant.smatrix(fsys, energy) in the script. Does it mean mumps is used in default? If yes, this explain why the default and mumps packages have the same speed in my simulation. I test your script in a window machines, SuperLU time is 2.2 times of mumps. In my window machine, the sparse installation does not have the umfpack installed. For the Ubuntu machine, the sparse method failed with the error: Traceback (most recent call last): File "/home/kayitsang2/Desktop/kwant ribbon/ribbon_testsolverV2.py", line 209, in <module> smatrix=ksparse.smatrix(fsys,energy) File "/usr/lib/python3/dist-packages/kwant/solvers/common.py", line 366, in smatrix flhs = self._factorized(linsys.lhs) File "/usr/lib/python3/dist-packages/kwant/solvers/sparse.py", line 105, in _factorized return factorized(a) File "/usr/lib/python3/dist-packages/kwant/solvers/sparse.py", line 83, in factorized umf.numeric(A) File "/usr/lib/python3/dist-packages/scipy/sparse/linalg/dsolve/umfpack/umfpack.py", line 450, in numeric umfStatus[status])) RuntimeError: <function umfpack_zi_numeric at 0x7f330c201158> failed with UMFPACK_ERROR_out_of_memory I can run the sparse section in ubuntu by reducing the size to 100X100, but the time used is 13 times of the mumps time. It seems the sparse package is restricted in memory and low in performance. I have installed the Ubuntu version using the command sudo apt-add-repository -s ppa:kwant-project/ppa sudo apt-get update sudo apt-get install python3-kwant python-kwant-doc if default is mumps, my problem is solved. Perhaps you also like to know that the time required for 200x200 is 123 second for mumps in the Ubuntu. The mumps time on window is 304s and the superlu time is 686s With regards, KS On 18/5/2017, 3:56 PM, "anton.akhmerov@gmail.com on behalf of Anton Akhmerov" <anton.akhmerov@gmail.com on behalf of anton.akhmerov+kd@gmail.com> wrote: Dear KS, On a linux workstation the ratio of runtimes is around 2.3 in favor of MUMPS, so the problem seems to pertain to your installation or to the windows installations in general. Can anyone using Kwant on windows try the attached script? (I reduced the number of points so that it doesn't take 1/2 hour and removed some parts not crucial to the benchmark) Best, Anton On Tue, May 16, 2017 at 7:18 PM, Prof. CHAN Kwok Sum <apkschan@cityu.edu.hk> wrote: > Dear Anton, > To confirm my observation, I remove the non-essential part of the script to test the mumps solver and default solver. The system is a zigzag graphene ribbon, with 200X200 unit cells, there are 4 atoms in one unit cell so there are 1.6x10^5 atoms. The OS is window 10, kwant is 1.2.2 and python is 3.4.4. the time for calculating 20 smatrix is around 30 minutes for both default solver and mumps. Please find attached the script and the runtime. The first runtime is for mumps. The second runtime is for default. The kwant was installed with the package prepared by Gohlke. > For the sparse solver the time is longer, 40 minutes but there is no umfpack installed in the window system. So I do not show the runtime as it may not be accurate. > KS > > On 16/5/2017, 5:51 PM, "anton.akhmerov@gmail.com on behalf of Anton Akhmerov" <anton.akhmerov@gmail.com on behalf of anton.akhmerov+kd@gmail.com> wrote: > > Dear KS, > > > I have compared the performance of default, mumps solvers in windows and > > Ubuntu machines. Their performance is very similar without significant gain > > when using mumps. Do I miss anything in the installation? For example, not > > installing the mumps. Similar performance is found for the sparse solver. > > The system I considered has 10^5 to 6X10^5 sites and machine has a quad core > > i7 cpu with 16G RAM. My installation procedures are those found in the > > website for pre complled packages. > > This is contrary to our experience. Can you please provide more details: > * How did you install Kwant in both cases? > * What specific script did you run? > * What were the execution times that you observed? > > Thanks, > Anton > > > KS > > > > > > > > Disclaimer: This email (including any attachments) is for the use of the > > intended recipient only and may contain confidential information and/or > > copyright material. If you are not the intended recipient, please notify the > > sender immediately and delete this email and all copies from your system. > > Any unauthorized use, disclosure, reproduction, copying, distribution, or > > other form of unauthorized dissemination of the contents is expressly > > prohibited. > > > > > Disclaimer: This email (including any attachments) is for the use of the intended recipient only and may contain confidential information and/or copyright material. If you are not the intended recipient, please notify the sender immediately and delete this email and all copies from your system. Any unauthorized use, disclosure, reproduction, copying, distribution, or other form of unauthorized dissemination of the contents is expressly prohibited. Disclaimer: This email (including any attachments) is for the use of the intended recipient only and may contain confidential information and/or copyright material. If you are not the intended recipient, please notify the sender immediately and delete this email and all copies from your system. Any unauthorized use, disclosure, reproduction, copying, distribution, or other form of unauthorized dissemination of the contents is expressly prohibited.
Prof. CHAN Kwok Sum wrote:
Does it mean mumps is used in default? If yes, this explain why the default and mumps packages have the same speed in my simulation.
kwant.smatrix is the same as kwant.solvers.default.smatrix [1]. This “default” solver combines the best available solver functionality [2]. So far the best solver is the one that uses MUMPS. [1] https://kwant-project.org/doc/1/reference/kwant#from-kwant-solvers-default [2] https://kwant-project.org/doc/1/reference/kwant.solvers#module-kwant.solvers...
participants (4)
-
Anton Akhmerov -
Christoph Groth -
Michael Wimmer -
Prof. CHAN Kwok Sum