Mailserver

5 min read

Hi, performance

The latest LXF arrived yesterday, thanks. In it I saw a note about Linux and high-performance computing, which finally prompted me to extract my digit. A few years ago, I wondered about whether to suggest a series on the subject of parallel computation but never managed to find a round tuit.

If you run with this, I suggest that a three-month series be appropriate, in the style of numerous other topics LXF has covered. The last time I used Nvidia’s CUDA for anything non-trivial was 12 years ago and things have changed markedly since then. More politically correct, though, may be to use OpenCL for tutorial purposes as it is free software and does not require any dedicated hardware. I have never used OpenCL, unfortunately.

Neil says… It’s interesting to see Intel’s new Arc GPUs scoring so well at OpenCL workloads. It seems like an obvious win for Intel if it could try to win compute in the server room, but it does look as though Nvidia is going to rule the roost for a long while yet, and you can’t deny that it deserves that with all the work it’s put into CUDA development.

Fortranate

I just read Mike Bedford’s excellent article in LXF307. I appreciated the mention that FORTRAN routines are often used under the hood in other languages. I’ll enjoy informing my graduate students, who are devoted to Python, that they’re sometimes using FORTRAN code.

If you want low-cost non-CUDA GPU acceleration, consider the Intel Arc options.

I have two additions to the history, in case of interest. Besides the IBM 704 platform that Mike Bedford mentions, DEC’s PDP-12 was another computer that had a FORTRAN II compiler, in which I wrote a suite of statistics programs while in grad school long ago. And my understanding is that MATLAB was originally written in FORTRAN, then rewritten in C. I’ve assumed that that’s why MATLAB starts vectors at location 1 (like FORTRAN) instead of location 0 (like C). I imagine that the rewrite couldn’t afford to break existing code.

MATLAB can still call compiled FORTRAN functions, which run faster than compiled MATLAB or compiled C linked to the same MATLAB front-end, according to a colleague who tested them.

Neil says… It’s fascinating to hear about everyone’s experiences over the years. Keep writing in with them!

Dualism

I’ve seen you say a number of times that using two separate hard drives is better, because “it works better”, without elaborating on exactly why. I’m currently using a laptop with Windows Ubuntu dual-booting on a single SSD, and it works just fine.

Neil says… An absolutely great question and you’re right, it’s something we just throw out