Perhaps this would be the book you need? http://www.kickstarter.com/projects/nickblack/the-finest-machine :D

I advise reading "My Other Computer is Your GPU" by myself and Jason Rodzik from earlier this year:

http://dank.qemfd.net/dankwiki/images/d/d2/Cubar2010.pdf

It covers these topics, and many more. :D

FORTRAN is still used so heavily because a) the architectures used for HPC are some of the rarest, most singularly-purposed machines on earth -- your Crays, your J-Machines, your Cell BE"s. These are not the commodity architectures for which gcc is designed; in many cases, gcc support might not even exist. In any case, it's certainly not going to have the intense and often, frankly, bizarre optimizations necessary to properly make use of these machines. Does python even allow one to manage cache prefetching, SIMD, the floating point model, etc? b) FORTRAN's easily parallelized by compilers -- even moreso than C with OpenMPI, etc -- due to the simplicity of its loop constructs and a rich heritage of automatically parallelizing these constructs (look at Ken Kennedy's work etc. google for DOALL and DOACROSS). For HPC, you're often working with massively parallel architectures, often SMID-heavy. Check out the Banerjee inequality sometime if you want to see something cool. c) Vector notation in FORTRAN is easily compiled into vector registers and operations. Python slices could probably be amenable to this, but Python isn't really ... d) Array-oriented. Everything in the microarchitecture world is designed around fast access to and operations on arrays. If you're not working in the array model, it's gonna be hard to compile it in a fashion that one can make use of architectural features designed for high performance. Language research is ongoing that seeks to address these issues. Python doesn't get there.