> Double descent with overparameterization is exhibited in "classical settings" too and mentioned in older books.
I’m curious for references or citations to this. When I was going over double descent I tried to find citations like this (just in a couple places like ML/stats textbooks).
> Because threads are so sensitive to initial conditions, it ends up skewing an entire discussion.
I haven't thought about this before on HN, but it makes a lot of sense. I'm curious if you or others have written about this intuition and your experience with it -- I'd like to understand it more.
Cool! I did this a while ago for some CPU-bound stuff, and ran into a bunch of performance bottlenecks.
Some things that helped me scale ptrace-interception up:
- SECCOMP_BPF filter (getting these right matters a lot)
- moving all of your intercept work to a single side (enter or exit)
- ensure affinity between the traced and tracing processes
- nuke vdso
- remove vdso from the aux vector (otherwise good libc's will find it again)
At the end of the day unfortunately the better solution would have been to write kernel support for what I wanted to do, but it's a fun exercise in learning about system calls.
Also recommend this as a follow-on. I think it's an important point, and also well presented. The anecdotes make it memorable but it's concise where it matters.
After I read this last year it changed the way I interacted and observed things around me for a while. Poking/prodding more, and asking more "what/how" questions.
Syscall interception works for _every_ program, it's just a matter of doing it correctly.
VDSO is a small set of (3) calls which are not syscalls but direct calls (for speed/efficiency). Our goal is to remove this functionality to force libs to call through the (slower) syscall route instead.
I mention in another comment how EHDR censoring is needed for robust VDSO removal.
I've not run into a libc where censoring EHDR breaks time calls (i.e. it doesn't fallback to syscalls) but possibly golang has this.
In this case it's straightforward to setup a fake VDSO and then instead of EHDR censoring you just replace it with your fake VDSO address and you're golden!
Ironically, because the folks working on containers/VMs are _really_ good at what they do, time access calls in particular have been really optimized (they get called a lot). This makes it very hard to intercept time calls at this layer! e.g. KVM and LXC both essentially hand time calls straight to the host.
This means time intercepts at the VM/container layer need fundamental support (I mentioned affine time transformation in the linux kernel in another comment) which doesn't work for people who need to deploy on current hosted container.
Nice Implementation! I built something similar a while ago to warp time in video games (for training reinforcement learning agents).
Some issues off the top of my head (that I ran into):
VDSO censoring is a lot harder than just symbol overriding, it has to actually be removed from the aux vector (third thing on the process stack when the process launches after arguments and environment variables. The EHDR entry is what you need to remove.
Some libc's (notably glibc) are really good at finding/using EHDR even if you do that symbol overriding, so dumping EHDR is the most assured way of making sure it's gone.
ptrace overhead is HUGE -- because you're debugging a userspace program with another program every time call now results in 4 context switches (to/from your debugging program at every time call entry/exit), even pinning both to the same CPU this is not fast.
This is where my least favorite part of the linux kernel comes in handy: SECCOMP-BPF. Instead of firing _every_ syscall, you can write a syscall packet filter rules list that only matches certain time-based syscalls with certain arguments. This greatly improves the performance (but for me, still not fast enough to play video games live).
At the end of the day I ended up reviving a >10 year old patch someone sent to the linux kernel to add these parameters (time offset and time warp) to thread structs and do the warping in the kernel (much faster -- dont pay the context overhead, etc). Sadly even this didn't work because our end application needed to run on multiple clouds in docker, and we'd need to have access to the host kernel to do these operations.
I'd like to have an affine time warp as part of the cgroups, and then maybe extend it through runc so anyone can run time-warped docker containers, but maybe that's wishful thinking.
Overall I think this is great work, and super happy you posted it. I'd love to chat about it sometime.
(P.S. most ironic to me was my version of this was called 'timelord' :)
Good intuitions. FCC regulations restrict EIRP[0], which is essentially "what is the signal strength in the strongest direction". It's possible that increasing directionality of your antenna puts you over the EIRP limit (given how routers already try to get _really_ close to it to improve performance).
It looks like we're moving towards a solution (in a slow, jerky, shambling way) with replacing hardcoded boardfiles with device trees.
If RISC-V does this well out of the gate (like with IBM's old school Open Firmware) and has great shiny device tree support, I think we might end up like x86 -- where a single thumbdrive can boot almost any x86 machine under the sun.
Followed the original RepRap project as a college student. A bunch of us students tried to do a group build of the RepRap Huxleys. It was way too much work to try off the bat in order to get a working printer.
After a run-in with a RepRap Mondo (it's big but it has a lot of issues, the design ended up not being that popular) bits-for-bytes RapMan printers (purchased by the industrial engineering department) and MakerBots (some folks at a local TechShop had them), we eventually got an Ultimaker kit.
It was amazing. Fast prints, cheap (compared to some), good-quality software, the assembly was straightforward and fast, instructions were thorough (compared to many other printers we've worked with), and print quality was good with almost no tuning.
Now graduated, I own an Ultimaker and whole-heartedly recommend it. Not only is it a good printer by itself, it has quite a clever community of inventors around it, and 3D-printed parts and improvements to Ultimakers appear with shocking regularity. (As opposed to makerbot, which is more of a professional printer in that less folks tend to modify them).
Cura, the software used to control ultimaker, has come leaps and bounds farther than what it started as. It's now easily mistaken for professional paid software, and it's easy enough that even someone with no experience can hop on and start printing.
Also shout-out to octoprint, which uses a Raspberry Pi to control the printer (3-hours of printing an awesome part rock, but not having your laptop tied to the printer for that long).
I’m curious for references or citations to this. When I was going over double descent I tried to find citations like this (just in a couple places like ML/stats textbooks).