BTW: So long as this is my pinned tweet, I'm happy to chat via DM about introductory systems-y things (like compilers, OSs, language runtimes, concurrency, etc.) or early-career stuff at BigCos in those kind of tracks. No obligation to know me IRL or on here or anything.

Anyone have any recommendations for a GPU version of Shen and Lipasti's "Modern Processor Design"? I'm ~90% certain such a thing doesn't exist. Maybe the closest approximation is the Synthesis Lectures "General-Purpose Graphics Processor Architectures"?.

Just had a Zoom call with a solar contractor and 5 minutes in he was like "Look, I'll be straight with you: we're booked through the end of the year, so you're not gonna go with us." And so we spent the rest of the call talking about how to pick one of his competitors.

I think my answer is physics -- the min fixed cost is big for disk and small for RAM, and transactional-ish strategies impose per-access penalties. So if you assume that the average transaction is small, fixed costs dominate for disk and per-access costs dominate for RAM.

To ponder: transactional interfaces have ~won for persistent storage (if you're unopinionated, you'll get a SQL database by default), but ~lost for transient storage (i.e. RAM). Why?.- Physics (i.e. one medium is better for one strategy).- Norms.- Research investment.- Other?.

I feel like every day I see some post that's like "LLMs can now do research-level mathematics", and then I try to ask it to do something I'd expect to be solvable by like a competent "intro to OS"-level student and it just falls over.

Has anybody done a writeup on using AI tooling for low-level/systems-y stuff? I'm having much more trouble getting reasonable C code out of it in a perf-sensitive domain than I think is common for e.g. more business-logic-y things.

Highly recommend it btw for any Bay Area people -- Dandelion Chocolate, in SF.

Wife and I finally got around to going to the makeup chocolate-making class that we had originally scheduled for April 2020.

I bet the Jeremy Renner app would have been popping off about this.

I wish there were some website I could go to that would feed me constant updates, riffs, etc., on my favorite topic: the details of the infidelities of B2B SAAS executives.

Remembering when Steve Ballmer gave a talk at my college a little while after Vista came out and in answering a question he was like "To a first approximation, Microsoft will live or die by the quality of the last Windows release. <pause> I mean, the last two Windows releases.".

(I would love to find out I'm wrong about this and there's some good strategy that avoids all these issues).

Today the choice is between "one big commit per feature" which is hard to review, "lots of small commits" which hurts reading history and obscures motivation, and "keep this info in some out-of-tree review system", which means you always need to look in two places for context.

A workflow I wish git supported better was subdividing commits into subgroups without losing linear history. I.e. "in this chunk of 5 commits, we're accomplishing X high-level task", and then each commit in that chunk could be as small as feasible.

(Norman Rockwell Four Freedoms guy): We should upzone Sunnyvale.

becomes almost trivial to implement if we assume that both processes periodically call a quiesce() function -- just pass off ownership in sequence. (2/2).

My answer: the hardware has a laxer API, in that it has weaker liveness requirements (if it sets a signal in the shared state, it "knows" that whoevers listening on the other side will acknowledge and respond to that signal). By comparison, software mutual exclusion (1/?).

(It relies fundamentally on the guarantee provided by the coherence implementation that there's one winner on writes to the turn variable, which more or less reduces to solving mutual exclusion; it just happens at the level of the coherence protocol).

A similar puzzle in memory models: Peterson's algorithm implements two-process flicker-free mutual exclusion of writes to shared state, but it's implemented on hardware units communicating via shared state (voltages on wires) that don't natively provide that abstraction. How?.

Anyone know of any experimental data on:.- (For GC'd languages) Gap between last access of an object and the time it becomes unreachable.- (For manual-MM languages) Gap between last access and the time it becomes free.?.