[CP2K-user] [CP2K:16769] Re: Question about COSMA

[Ole Schütt]

Hi Hongyang,

the COSMA library uses a novel algorithm that minimizes communication by 
using more memory. This trade-off usually leads to better overall 
performance. However, it might be that in your particular case 
MKL/Scalapack is actually the better choice. You can easily switch between 
the two via the input:

https://manual.cp2k.org/trunk/CP2K_INPUT/GLOBAL/FM.html#TYPE_OF_MATRIX_MULTIPLICATION

Alternatively, you can also reach out to the COSMA developers directly by 
opening a Github issue: https://github.com/eth-cscs/COSMA/issues

-Ole

On Thursday, March 24, 2022 at 10:25:32 PM UTC+1 ma455... at gmail.com wrote:

> Hi Ole,
>
> Thanks for your response. But I actually didn't use gsl when I compiled 
> cp2k. My compiled file is attached. Please have a look.
>
> Best,
> Hongyang
>
> 在2022年3月25日星期五 UTC+11 08:20:31<Ole Schütt> 写道：
>
>> Hi Hongyang,
>>
>> we have been puzzled by this mysterious slowdown as well 
>> <https://github.com/cp2k/cp2k/issues/1826>. Yesterday we finally had a 
>> breakthrough: You have to remove `-lgslcblas` from your ARCH file to 
>> ensure that the proper BLAS library gets linked.
>>
>> -Ole
>>
>> On Thursday, March 24, 2022 at 10:51:19 AM UTC+1 ma455... at gmail.com 
>> wrote:
>>
>>> Hi,
>>>
>>> I'm just wondering are there any advantages of using COSMA instead 
>>> of SCALAPACK for matrix multiplication? I found the speed of using COSMA is 
>>> slower than SCALAPACK in SCF calculation.
>>>
>>> Thanks,
>>> Hongyang
>>>
>>

-- 
You received this message because you are subscribed to the Google Groups "cp2k" group.
To unsubscribe from this group and stop receiving emails from it, send an email to cp2k+unsubscribe at googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/6f5852f0-b16a-436b-9625-b97aaec16d26n%40googlegroups.com.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20220324/983ae240/attachment-0001.htm>