R devel - Mar 2007 - Incorrect matrix of spearman correlations .... in 64-bit Linux ... (PR#9568)

Full_Name: Vladimir Obolonkin
Version: tested in 2.0 to 2.4.1
OS: linux, win, mac
Submission from: (NULL) (202.14.96.194)

{{ Subject shortened manually -- to pass anti-spam filters 

   Original-subject: Incorrect matrix of spearman correlations working \
     large (24000 by 425 and 78 by 425 data frames)  in 64-bit Linux machines;\
      the same code gives correct results in 32-bits Win and Mac (PR#9568)
}}

cc_s<-cor(phenos,vec,method="spearman",use="pairwise.complete.obs")

this is a line copied from real script producing different results in 64
bit/Linux/R (wrong) and Mac&Win (correct)

The script was implemented on 4 machines with Linux 64, Suse 9-10, in 5 variants
of R -- versions from 2.3 to 2.4.1, compiled with few different settings of
optimization and 2 versions of compilers. In all cases the results of spearman
correlation were identical, but wrong.

The same script was started up to 10 times in Win32 on Intel and Linux32 on Mac
with the Rs from 2.3 to 2.4.1 -- in this set of cases the results were identical
and correct.

'phenos' and 'vec' are data frames of 425x78 and 425x24128
respectively, all
numeric variables. The 'phenos' has moderate number of NAs in some
columnes.

The problem dissapeared when trying to reduce the size of matrices (by selection
of rows and/or columns) and/or when simulating the data with random generators.

vobolonkin at lic.co.nz wrote:
> Full_Name: Vladimir Obolonkin
> Version: tested in 2.0 to 2.4.1
> OS: linux, win, mac
> Submission from: (NULL) (202.14.96.194)
>
> {{ Subject shortened manually -- to pass anti-spam filters 
>
>    Original-subject: Incorrect matrix of spearman correlations working \
>      large (24000 by 425 and 78 by 425 data frames)  in 64-bit Linux
machines;\
>       the same code gives correct results in 32-bits Win and Mac (PR#9568)
> }}
>
>
cc_s<-cor(phenos,vec,method="spearman",use="pairwise.complete.obs")
>
> this is a line copied from real script producing different results in 64
> bit/Linux/R (wrong) and Mac&Win (correct)
>
> The script was implemented on 4 machines with Linux 64, Suse 9-10, in 5
variants
> of R -- versions from 2.3 to 2.4.1, compiled with few different settings of
> optimization and 2 versions of compilers. In all cases the results of
spearman
> correlation were identical, but wrong.
>
> The same script was started up to 10 times in Win32 on Intel and Linux32 on
Mac
> with the Rs from 2.3 to 2.4.1 -- in this set of cases the results were
identical
> and correct.
>
> 'phenos' and 'vec' are data frames of 425x78 and 425x24128
respectively, all
> numeric variables. The 'phenos' has moderate number of NAs in some
columnes.
>
> The problem dissapeared when trying to reduce the size of matrices (by
selection
> of rows and/or columns) and/or when simulating the data with random
generators.
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>   
I don't see anything to reproduce here, so how are we supposed to get a 
handle on the issue?