Correlating profiles across different machines usually involves a lot of painful window-switching.
Here’s a quick demo of doing it in Perfetto instead: two separate profiles, aligned on one timeline in a few clicks.
The trace merging feature in Perfetto v58 has a lot of cool applications: view independent traces from different VMs/devices, client-server interactions, and most topically, multi-rank and multi-machine PyTorch profiles.
Cooking up a blog post about all of this, coming soon.
Perfetto v58 is out!
We shipped a bunch of things: persistent CLI analysis sessions for agents, zstd compression, better memory tools and more.
Biggest feature though is trace merging: multiple traces on one timeline, including across VMs and devices.
https://t.co/HtElsouquc
Perfetto v58 is out!
We shipped a bunch of things: persistent CLI analysis sessions for agents, zstd compression, better memory tools and more.
Biggest feature though is trace merging: multiple traces on one timeline, including across VMs and devices.
https://t.co/HtElsouquc
The staff engineering post I wrote a couple of weeks ago is currently #1 on Hacker News, which is a nice surprise.
If you missed it: https://t.co/sfFNKKdNd4
The staff engineering post I wrote a couple of weeks ago is currently #1 on Hacker News, which is a nice surprise.
If you missed it: https://t.co/sfFNKKdNd4
@jdxcode Excellent timing! I'm just polishing the build profiling tool for release and was using clap; was hoping you'd publish usage-rs so I could try it out before publishing.
@jdxcode Sure those questions are very important and worth spending time on.
But does that mean you're finding AI at architectural decisions good enough? Just trying to work out if I'm genuinely missing something as I don't find myself able to take my hands off the wheel in that domain.
@softminus@the_aiju For pure CPU bound work, flamegraphs and staring at assembly is the way. But when your system gets complex enough, a lot of the problems start becoming "I'm waiting on X" where X is a mutex, a socket, a file handle etc. And at that point you end up needing tracing!
@jdxcode Yes and no; it's one less round trip but I agree it's more confusing because most CLIs don't do it so you need to remember the tools which do.
@ryiacy I've definitely noticed this while building Perfetto as well. Though I'd push back a bit; I do think it's still useful to have some simpler entrypoint both for orienting the AI and for keeping the user somewhat feeling like they know what's going on.
@valigo@oxcrows IMO this is *the* biggest thing which WASM is missing. There's so many cool things you can do if only you could just use mmap in WASM...
5/5 Caveat: the samples of Opus 5 is small; mainly because I don't particularly like using it (mainly been using Sol but I ran out of usage for the week!)
1/5 I've been on Opus 5 for a week and noticed the code comments stopped using em-dashes. It's all double hypens (--) now.
Decided to do some analysis to confirm I wasn't fooling myself...
@PremiumGoblin Nice to see another new L7 SWE on here :)
What have you found as the most interesting project of the ones you've worked on and why? Interested to hear both from a technical perspective and an organizational one!
@YangWang92 No worries, I understand! I'm always interested in workflows like this though because it helps us understand better the pain points in how things work today!
> Thanks for building Perfetto!
You're very welcome, it's been a pleasure!