[petsc-users] ParMETIS_V3_PartGeomKway

[Gerard Gorman]

Matthew Knepley emailed the following on 16/02/12 13:46:
> On Thu, Feb 16, 2012 at 7:43 AM, Gerard Gorman
> <g.gorman at imperial.ac.uk <mailto:g.gorman at imperial.ac.uk>> wrote:
>
> Matthew Knepley emailed the following on 16/02/12 13:29:
>> On Thu, Feb 16, 2012 at 3:38 AM, Mohammad Mirzadeh
>> <mirzadeh at gmail.com <mailto:mirzadeh at gmail.com>
>> <mailto:mirzadeh at gmail.com <mailto:mirzadeh at gmail.com>>> wrote:
>>
>> Hi guys,
>>
>> I'm wondering if there is any implementation for
>> ParMETIS_V3_PartGeomKway()? All I can find is the implementation
>> for ParMETIS_V3_PartKway and I'm wondering if including the vertex
>> positions could help me get a better partitioning?
>>
>>
>> As far as communication goes, it will not help.
>>
>>
>> Also I have a general question. Is minimizing number of edge cuts
>> essentially the same as minimizing communication? I understand that
>> they are related, but communication is proportional to the number
>> of ghost points which is not exactly equal (its actually less than)
>> to the number of edge cuts. So then is it possible that this could
>> actually result in larger number of ghost points, at least for some
>> of processors?
>>
>>
>> It of course depends on your problem and the graph you draw, but
>> you can always draw a graph where the edge cut is exactly your
>> communication.
>>
>>
>
>
> It might also be interesting to look at Zoltan's hypergraph
> partitioning which can better balance communications -
> http://www.cs.sandia.gov/zoltan/dev_html/dev_phg.html
>
>
> I would caution you that one thing people are rarely careful about is
> quantifying the effect of latency vs communication volume. I would
> bet you $10 that the lion's share of slow down is due to load
> imbalance/latency rather than communication volume (since comm links
> are so big).
>
> Matt
>

I believe you'd be correct in the vast majority of cases. Anyhow, as
Daniel Bernstein says, "Profile. Don't speculate."

Cheers
Gerard