Seattle, WA
Joined February 2016
- Tweets243
- Following114
- Followers112
- Likes157
Pinned Tweet
New abomination: a tool that spawns a shell session where there are no consequences
It uses overlayfs to redirect all writes to an overlay upperdir instead of messing up the real filesystem. Ditch all changes just by exiting the shell session: github.com/vikmik/scratch
So one of my programs runs ~60% slower (from ~18s to ~31s) after removing an unused function, due to Intel's JCC bug mitigation - running it on a core i7-10710U
Outlier for sure, but damn that's far from Intel's advertised average 0-4% performance hit
Source code for the curious (includes binaries and assembly). Seems to repro with clang 11/12/13, though on clang11 the results are reversed.
github.com/vikmik/jcc-perf-h…
Given that compiler workarounds (branches-within-32B-boundaries) aren't generally used, the real collective cost of this would be interesting to know.
Looking at the Linux kernel, Debian glibc & Mathematica, affected jumps are everywhere
Makes it hard to trust benchmarks too!
On the Skylakes that didn't get their LSD disabled, are there documented corner cases of JCC erratum mitigation not behaving as it should?
This uops.info snippet with offset 50 bit.ly/3fJnTJx has a suspiciously high count of DSB+LSD when I actually run it
With cmp+jg macrofused and crossing a 32-byte boundary, I would not expect the DSB & LSD to kick in.
However they definitely do when the instructions "cleanly" cross a 64-byte boundary (op before, jcc after), contradicting (I think) Intel's docs?
(case 6 below)
Admittedly I'm pretty outside of my comfort zone, and also have half a functioning brain right now. Wondering if I can invoke the mighty @trav_downs or @dendibakh ? :)
Here's the nasm snippet I'm using to repro: pastebin.com/Kt6qKTbF
Victor Michel retweeted
Mike Freedman's proof of the 4d Poincare conjecture is one of the most daring, inscrutable results in math history. Now, a new book is trying to save it from being forgotten. I cover this decades-long drama in my new piece for @QuantaMagazine quantamagazine.org/new-math-…
Victor Michel retweeted
If you like kernel deep dives, you'll like this. @vic_mic_ did a followup for my demand paging piece, with a performance focus, covering a LOT that I missed.
My favorite part: the clever page-table fu that triggers pathological kernel allocation behavior via a top-down scan. 🥋
New blog post: @vic_mic_ takes Prodfiler down into the depths of the Linux memory management system ... to examine "The OOMsday device".
prodfiler.com/blog/the-oomsd…
Victor Michel retweeted
New blog post: @vic_mic_ takes Prodfiler down into the depths of the Linux memory management system ... to examine "The OOMsday device".
prodfiler.com/blog/the-oomsd…
Victor Michel retweeted
That piece from John Conway's "A course in functional analysis" opened my eyes - way too late - back when I was a math student
"The human mind resists abstraction unless it first sees the need to abstract" is a very good way to put it
Victor Michel retweeted
Come check out what we've been working on! C/C++, Rust, Go, Perl, PHP, Python, Java/JVM and Ruby. Kubernetes/docker/whatever. No need for on-host debug symbols or service restarts, and an overhead that's low enough to leave it running all the time, everywhere =)
It is alive! Introducing Prodfiler, the world's first frictionless whole-fleet whole-system continuous profiler: prodfiler.com/blog/introduci… -- profile all your code, everywhere, all the time.
Try it today :-)
Victor Michel retweeted
New "Painting with Maths" (loooong) video in Youtube:
youtube.com/watch?v=8--5LwHR…
Made in Shadertoy (raymarched SDF): shadertoy.com/view/WsSBzh
#Mathematics #SDFs #raymarching #procedural #shader
Victor Michel retweeted
Anecdote: For 15 years floppy disk copy protection was fun cat+mouse hackery (Apple2, C64, Amiga; I was friends with the XCopy guys+I did 'Cyclone' hw copier) with lots of clever/complex tricks (sector sync, errors, long tracks, fat tracks, short tracks, 'weak bits' etc etc)...
