[DFTB-Plus-User] Problem of parallel computing with openmpi

[yokoi-mp.pse]

Dear Dr. Bálint,

Thank you very much for your advice.

Actually I use a normal 1 Gb networking switch, rather than an 
infiniband switch, so it may be one of possible problems.
I will try a larger number of atoms varying cores and nodes.

Best regards,
Tatsu

On 2017/5/11 23:23, Bálint Aradi wrote:
> Dear Tatsu,
>
>> When I perform a MD calculation on 4 cores in one node, computational
>> time effectively decreases compared to a single-core calculation.
>> However, when the same MD calculation is done on 8 cores of two nodes,
>> its computational time become 2-3 times longer than the case of 4 cores
>> in one node.
>>
>> I think that the calculated simulation box contains relatively a large
>> number of atoms (496 Si atoms) and the number of nodes is just two, and
>> thus communication overhead is not a cause. Would you give me advice
>> about possible causes of this problem?
>>
> It depends pretty much on the interconnect between your nodes. I could
> get 1000 atoms scaling up to 4 nodes, when using nodes with 8 cores each
> and connected by infiniband. Your system is rather small (you have
> roughly half of the orbitals I used), but I still would accept scaling
> to 2 nodes, provided you have a low latency interconnect.
>
>    Best regards,
>
>    Bálint
>
>
>
> _______________________________________________
> DFTB-Plus-User mailing list
> DFTB-Plus-User at mailman.zfn.uni-bremen.de
> https://mailman.zfn.uni-bremen.de/cgi-bin/mailman/listinfo/dftb-plus-user

-- 
----------------------------------
横井　達矢
名古屋大学　大学院工学研究科　物質科学専攻

〒464-8603 名古屋市千種区不老町
名古屋大学工学部5号館633室
Tel/Fax: 052-789-3365
e-mail: yokoi at mp.pse.nagoya-u.ac.jp
----------------------------------

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.zfn.uni-bremen.de/pipermail/dftb-plus-user/attachments/20170512/c5129138/attachment.htm>