Intel® Fortran Compiler
Build applications that can scale for the future with optimized code designed for Intel® Xeon® and compatible processors.

Problem with big arrays?

Cayo_G_
Novice
1,489 Views

Hello. I have a code that was running just fine. Then, I increased the size of my arrays largely and it sometimes doesn't compiles and I tried with some flags that compiled but crashed with errors I don't understand. 

I had some arrays of size about 3000, that I changed to 22554 ( and also squared matrices of this size). I have only my main code and a module, in which I create most of my global variables. 

I was originally compiling just with: 

ifort -c global_param.f90; ifort -c dyn.f90; ifort global_param.o dyn.o -check bounds

And for that I got the error during compiling:

global_param.o: In function `global_param_mp_hamiltonian_':
global_param.f90:(.text+0x145): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot1_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x14e): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot1_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x173): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot1_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x1d9): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot1_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x1e3): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot1_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x212): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot1_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x22c): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot2_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x236): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot2_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x25a): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot2_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x2c0): relocation truncated to fit: R_X86_64_32S against symbol `global_param_mp_pot2_' defined in COMMON section in global_param.o
global_param.f90:(.text+0x2ca): additional relocation overflows omitted from the output

The I searched and found out that this could be due to the size of my arrays, and that I should compile with the -fpic and -mcmodel flags. So, I tried:

ifort -c global_param.f90 -fpic -mcmodel=large ;ifort -c dyn.f90 -fpic -mcmodel=large ; ifort -fpic -mcmodel=large global_param.o dyn.o -check bounds

And then I simply got the error

Segmentation fault

 

I don't know what to do. Can someone please help me?

Many thanks in advance,

 

Cayo

0 Kudos
1 Solution
jimdempseyatthecove
Honored Contributor III
1,489 Views

Make the large/huge arrays allocatable (then allocate at program start).

Jim Dempsey

View solution in original post

0 Kudos
7 Replies
jimdempseyatthecove
Honored Contributor III
1,490 Views

Make the large/huge arrays allocatable (then allocate at program start).

Jim Dempsey

0 Kudos
Cayo_G_
Novice
1,489 Views

Hello Jim Dempsey,

 

Thank you for your suggestion. It helped, but now I have a new problem. One of my variables is a matrix with dimensions of (67662,67662).

I did the allocation in the beginning of all my big arrays as you suggested, including this big matrix. The program now crashes when I try to give it a value. I simply do:

Ha(:,:)=0.d0

and the program crashes just there with the error:

forrtl: severe (174): SIGSEGV, segmentation fault occurred
Image              PC                Routine            Line        Source             
libifcoremt.so.5   000014B60BB792D6  for__signal_handl     Unknown  Unknown
libpthread-2.27.s  000014B609F06890  Unknown               Unknown  Unknown
dyn.exe            0000000000403707  Unknown               Unknown  Unknown
dyn.exe            000000000040610B  Unknown               Unknown  Unknown
dyn.exe            0000000000400D8E  Unknown               Unknown  Unknown
libc-2.27.so       000014B609B24B97  __libc_start_main     Unknown  Unknown
dyn.exe            0000000000400C7A  Unknown               Unknown  Unknown

What should I do now?

Again, many thanks in advance.

0 Kudos
Cayo_G_
Novice
1,489 Views

Oh, I just did Ha=0.d0 instead of Ha(:,:)=0.d0 and it worked. 

Thanks for the help

0 Kudos
jimdempseyatthecove
Honored Contributor III
1,489 Views

You should report this as a bug. Both syntaxes should work. (include your system info and IVF version information too).

They are going to ask you for a reproducer too. See if this breaks on your system, if so, use this as the reproducer

!  BigArray.f90 
program BigArray
    implicit none
    real(8), allocatable :: Ha(:,:)
    allocate(Ha(67662,67662))
    Ha(:,:) = 0.d0 ! whereas Ha = 0.d0 works
    print *, shape(Ha)
end program BigArray

Jim Dempsey

0 Kudos
Ron_Green
Moderator
1,489 Views

Remember you are constrained by the physical memory of your computer.  AND possibly stacksize limits for array temporaries. 

67662^2 is 4.578146e9  which is 4.6 giga-elements.  With single precision you need 18.4GB.  double again for Double and you're over 36GB

so in Linux you would hope the allocate would fail if you need too much memory. Nope, linux uses lazy allocation so you don't really touch that memory and allocate it until you first need it, which is the Ha(:,:) = 0.0_4 or 0.0_8 initialization.  If you don't have enough physical RAM and swap space BLAMMO!  seg fault on that first touch.

0 Kudos
Ron_Green
Moderator
1,489 Views

As a side note, static arrays are stored in a 2GB segment along with code.  Hence the initial compile errors when you exceeded 2GB in arrays.

0 Kudos
Cayo_G_
Novice
1,489 Views

Yes Ronald, you are correct. With all the others variables in my code, it is actually almost 70GB of ram. 

This 2GB thing you mentioned I saw somewhere else, and that why I included the -fpic and -mcmodel=large flags.

0 Kudos
Reply