Above the Fold
- Starcloud - from the NVIDIA blog. I think the main advantage to this concept is that in space, data canters are not currently subject to any local laws. The downside is that, if deployed as shown, the whole thing will melt in about 10 minutes as they forgot a heat dissipation system.
- How the brain moves from life to sleeping (and back) - Quanta Magazine
- Does Mars have a core at its center? This paper says it does.
November Office Hours
ORCD office hours are cancelled on Tuesday, November 11 and Thursday, November 27 due to Institute holidays.
What We’re Reading
- Ray Kurzweil and the singularity (MIT, '70) at MIT this month.
- Superintelligence angst
- New GPU supercomputers from AMD and NVIDIA. Lots of FP4 and FP6 flops, adieu FP64?
- Sample ChatGPT to get better jokes (and more).
Meet the Team: Lincoln Bryant, Lead Site Reliability Engineer
Lead Site Reliability Engineer Lincoln Bryant recently joined ORCD from the Enrico Fermi Institute at the University of Chicago where he led a team of systems engineers operating one of the largest dedicated computing facilities for the ATLAS experiment at the Large Hadron Collider. He was also highly active in the OSG Consortium, a national-scale fabric for High-Throughput Computing. Before working at the University of Chicago, Lincoln obtained a BS in Physics from the University of Cincinnati.
Outside of work, Lincoln enjoys reading speculative fiction novels and computing history books. He also dabbles in retrocomputing, game development, 3D modeling/printing, and functional programming.
Down a Rabbit Hole: Mathematica, SLURM, ChatGPT and Me
As reported in these pages earlier, I have developed a diffusion code to understand the transport of positrons from an exploding black hole to Earth. I wrote the code in Mathematica—as I learned it has a good solver called NDSolve—and its runtime is highly variable—hours to a week. For each parameter set, I need to run over 16 energies, so I set up bash scripts to submit jobs to SLURM, as one does.
The problem came when I found I needed to checkpoint: I had chunked the time integration into different periods to accommodate a variable source function, but even the runtime of a chunk varies by more than a factor of ten.
Checkpointing turned into a nightmare of confusion and misinformation: SLURM has the –signal feature that sends a signal to the bash script when the script approaches the maximum runtime. That works fine, and many use this. The problem came in the script’s communication to Mathematica: I could not make the trap function to prevent the signal from going to Mathematica, and then I found that any signal causes Mathematica to terminate. I asked ChatGPT about this and it told me that USR1 causes termination, but USR2 and CONT (continue) did not. I did some tests and proved to myself that USR2 causes the termination of all tasks. I pointed this out to ChatGPT and got, “Oh yes, very sorry, you are right.”
I then tried CONT and found that it has no effect on Mathematica when used interactively, but caused termination when used with SLURM—even more bizarre. I asked ChatCPT and, after churning for 20 minutes on the “Deep Research” setting, it told me Mathematica does not use signals, so I should switch to USR2. Grrrr…
Some have said, “ChatGPT works at the level of a first-year graduate student…” in coding. I really don’t agree with this, based on the experience above and other times I’ve used it. First, ChatGPT is not human and does not respond in a human way. It readily admits mistakes, which is good, but does not learn from them, even to the point of just repeating them. I start my sessions with, “I am trying to…” so it understands my larger goal, but forgets that after a few exchanges and focuses on the nuts and bolts questions, rather than suggesting a different approach. It never admits to not knowing something.
My bottom line is that I do not see ChatGPT as any sort of partner, but rather an idiosyncratic information source. For this project, I wish I had just read the details of Linux signalling, bash, and the SLURM doc to start and not bothered with ChatGPT.
-Peter
ORCD Fall Training Schedule
ORCD will be teaching one more class to members of the MIT community during the fall of 2025.
- Introduction to Engaging:
- Wednesday, November 12, 10 AM - 12PM
See the 2025 fall classes page on ORCD’s website for more information and signup instructions.