Hi, friends:

I'd like to get your advice on a mysterious problem. I have a code running ocean wave boundary layer simulation. And I'm running for 1-day realization, with time step about 0.1 sec, which I have about 864000 time steps to march in a big loop.

It runs okay on my linux PC but pretty slow. Then I turned to an SGI origion 300 with 4 CPU's in order to get results faster. Problem being it stops somewhere around 100,000 timesteps, without any error info.

I suspected it might be size of integer has a limit for loop index?

Please help me, thanks a lot

Wen

first of all if your code cannot take adventage of multiprocessor machine your code will not be faster.

secondly, it can be many things... checking sizes/ranges of floating/int variables is good to start with

then the fortran compiler is differnt i guess. if e.g. you run your code on your pc using gnu see if gnu is not available on SGI too.

then, write some log file on the fly to debug easier. success!

I have experienced situations when my codes run on linux but doesnt on SGI. In all cases there was a bug in the code. Check to see that you are not exceeding your array sizes.

Or this could be a memory issue, ie. if you use MPI you may need to define some memory parameters. The only way I can see this being the problem though is if you save a bunch of data each time-step, some of that in the memory. But it's more likely a bug, as all compilers are slightly different. A quick check could be to compile with -O3 for maximum optimization, this has sometimes helped me to discover a bug as the optimized code does something completely different than it is supposed to do.

-- Jarmo

Hi, dominik,Praveen and Jarmo:

Thanks a lot,

I did find a array bound exceeding bug after using both -O3 and -check_bounds option in compile, and fixed it. But the problem still persists. It's usual f77 code, no MPI things are used.

One thing I'm not familiar with SGI is that, when I run the code, the gr_top command will show me that 4 copies/threads of the code are running at the same time, and there is one at the status of sleep, when the code becomes idle. Anything wrong with the 4 CPU?

Thanks again,

wen

Thanks, dominik:

I found g77 on that machine too, and trying that out today.

Wen

Going back to Dominik's message:

How is the code parallelized if you are not using MPI? This might be the problem.

See if it runs with just one processor on the SGI.

-- Jarmo

Wen,

I guess you are using the 'apo' (autoparallelisation) switch to make your code parallel??

If it is the case, you might need to use some directives is a source code. The apo is not that clever as it sounds.

For finding another problems try ftncheck program for finding errors in the code. It's saving me hours every day. http://www.dsm.fordham.edu/~ftnchek/

Matej

Thanks, all your inputs add to my experience of debugging.

After careful debugging, I found the bug. I have a variable t accumulating time with very little step, like t=t+dt, while dt=1.e-4(sec) or so.. Turns out that when t is sufficiently large, say 1/3 day, then the rounding error of float addition will actually do nothing to t, which I used to jump out of loop, since t never changes beyond 1/3 day, it goes like a dead loop spinning. Ha, fancy that,

Wen

Don't you just love computers

But, congratulations for finding it. That didn't even take that long! I guess this is a good example why we should stick with double precision.

-- Jarmo