yade-mpi team mailing list archive
-
yade-mpi team
-
Mailing list archive
-
Message #00012
Re: Method on partitioning
-
To:
yade-mpi@xxxxxxxxxxxxxxxxxxx
-
From:
Bruno Chareyre <bruno.chareyre@xxxxxxxxxxxxxxx>
-
Date:
Wed, 28 Nov 2018 11:16:13 +0100
-
In-reply-to:
<CAPWBqP+YxDGB3krAEw0P7Ed6n5bM_8jR=jzvd34Oik2K5hV=+w@mail.gmail.com>
-
User-agent:
Mozilla/5.0 (X11; Linux x86_64; rv:60.0) Gecko/20100101 Thunderbird/60.2.1
Hi,
On 11/27/18 8:32 PM, Robert Caulk wrote:
Hey Deepak,
Feel free to add it! I should also point out that I just pushed 2
methods for automatic 3D domain decomposition. One of the methods uses
least squares to minimize the differences between subdomain x,y,z
lengths and [1].
Whenever "length" appears somewhere it is not a bad idea to define it as
max(x,y,z) instead of sqrt(x²+y²+z²). Just because this norm corresponds
to the axis-aligned bounding box norm.
If you ask it for an "odd" number of workers (domains) it splits some
of the domains to make use of remaining workers (you can also manually
control the decomposition so you could just decompose evenly along one
axis, maybe more suitable for "odd" worker requests). The other
decomposition method uses a sparsely generated random point cloud to
set an arbitrary number of randomly shaped subdomains [2]. Feel free
to test them out [3].
It seems we need to have a naïve decomposition for a start (fixed
cartesian grid, most likely) else it will be impossible to quantify
progress on this aspect in the hackaton.
B
References