Meetups/Infra/2025-07-07

Big turnout. Discussion of performance engineering. Lots of links: tools, methods, why / why / how.

Introductions

• [name] - [background]. [goals for meetup, or interests to explore]

• Loren -- platform engineer, want to talk about performance engineering, want to invite gpu perf specialists to attend in near future
• ... -- interested in reverse engineering, here to have fun & hang out
• Greg -- programmer by trade, interested in self-sovereign computing
• Yayoi -- programmer,
• Max -- do some home labbing
• Matty -- mathematician, haven't come for the meetup before, here for the vibes
• Alex -- ml, working on lil LED display
• Chris
• Gopal -- intersted in platform engineeer, have a home lab
• Kevin -- code for a living enjoy playing with data
• Doug -- like computers, no agenda
• Cloud -- 5mof.net, next on July 17th, 3rd thurs
• Eric -- mostly like hw stuff
• Wheezy -- here to learn, big on hacking closed source systems -- by contrast..
• Zacchae -- prb seen me around warning my open source wearables, do ml for a living

Topics sought

Commands opensnoop Dynamically trace and report when files are attempted to be opened or stat'ed by processes, in real-time e.g. snoop on open() and stat() system call part of the DTrace toolkit Use Cases: Log/Debug file-access events e.g. issues that occur sporadically or under specific conditions. Monitoring system calls in real-time for security auditing, troubleshooting, or optimizing file I/O operations. Print all file opens as they occur:

                 sudo opensnoop

Track all file opens by a process by name:

                 sudo opensnoop -n "process_name"

Track all file opens by a process by PID:

                 sudo opensnoop -p PID

hash A shell's command-hash table is a map of a binary's name to file-path location mechanism that shells use to specify the file-path location of the binary associated with a command List the file-path location of the binaries the shell associats with commands

                 hash

Rebuild / reenumerate the Hash Table

                 hash -r

scc https://github.com/boyter/scc Calculate language-relative code complexity

                 Complexity Algo uses: https://wikipedia.org/wiki/Cyclomatic_complexity
                 scc --by-file --sort complexity

Browser Privacy

     https://privacytests.org/    
   https://blog.cloudflare.com/announcing-encrypted-client-hello/  

PiRogue Open Source Incident response tool running on a Rasbberry Pi: https://pts-project.org

     https://github.com/PiRogueToolSuite    

Jobs: https://www.digitalrights.community/job-board OpenEmail

     https://www.open.email/    
     https://github.com/Open-Email/MailHTTPS-Protocol    

Censorship Circumvention Lantern

       https://lantern.io/faq      
       https://github.com/orgs/getlantern/repositories      

Psiphon free and open-sourceInternet censorship circumvention tool that uses a combination of secure communication and obfuscation technologies, such as a VPN, SSH, and a Web proxy.

       https://github.com/orgs/Psiphon-Labs/repositories      

• Decentralized Tech
  • Radicle (git self-host) desktop gui: https://github.com/radicle-dev/radicle-explorer
  • vs Forgejo (esperanto) https://forgejo.org/

• Performance / engineering
  • When is it relevant?

https://www.brendangregg.com/usemethod.html
https://mechanical-sympathy.blogspot.com/

Lesson or Demo

• Read aloud: clarify for meetup. We are taking notes in a riseup pad (or I am--help appreciated, and links). We have meeting notes posted to the wiki. noisebridge.net, search Infra, or Meetups/Infra. (the Infrastructure page has a disambiguation link.)
• Shell, web services, self-hosting, networking!

light-ms 186mi profiling with flamegraphs --

1. FreeBSD optimizations used by Netflix serving at 800Gb/s - Drew Gallatin - EuroBSDcon 2022

https://www.youtube.com/watch?v=36qZYL5RlgY

context -- depends on application, context, user -- sometimes hardware constraints

constrained, often on network architecture -- pref on order of network latency matters more than offline. depends hugely on domain. On web, 100ms can ok, or quite long

have a website

back of the envelope calcuations -- theoretical limits -- what's the total work, what's the max throughput of your hardware

flamegraphs

python[3] -X importtime ...
go -- pprof

rust -- had some issues using it

used tracy instead -- better way of getting summary statistics (among the worst installation instructions, has brew cask)
https://github.com/wolfpld/tracy

performance -- exploratory searching vs net effects

tracy -- require s

py-spy,  all manner of python /flamegraph/ search results.  pyflame.

Other performance triggers "it feels slow" "business critical -- and hot path" "tests slow" -- detect if data bad -- evidence of not "when it stops letting you do things"

"bottlenecks"

process, if you think this is relevant

- perf might be at issue - napkin math ? is this expected? what's the best possible throughput? amdhal's law, parallel - establish baseline

CPU vs GPU performance optimization

historical perspective -- frame budget, for gpus. in 1 frame, do a bunch of highly data parallel computation. - gpu sharing instruction cache

application of the intensity / latency tradeoff in your own work or projectcs

• have a web application -- static site, want thing to load fast all the time.

Linux Nginx Php Mysql stack

cache better?
cloudflare static hosting?

• building tools for llm/s. CRUD functions on databases. gemini models, develop mcps, for model agnostic

issue -- don't know the shape of compute.
tools you reach for now, to approximate the answer to that quickly?  no, idk
what I hate about programming on gpu

• doug -- doing inference, on retail data. lots of data, sparse matricies, very inefficient. always looking for ways to run thing faster. 20m work, better library 45s, on gpu 45s (each inferrence)

pytorch geometric
pytorch native 
jax (google)

performance tuning -- for cpu. justine tunney gave a talk -- ollama.cpp, llamafile

https://thume.ca/2023/01/02/one-machine-twitter/

https://accidentallyquadratic.tumblr.com/

reddit, performance engineering -- why not one computer?

https://nitter.poast.org/

https://github.com/mcint/bsky-firehose-term

Takeaways & for next week

• learned about tracy, telemetry pr0n on the screenshots
• triton -- open python gpu compiler
• good convo
• learned about gpus -- check out CloudFlare for client side issue. try in chrome.
• cyclomatic complexity tool looks p cool // like $ cloc (cound lines of code) but cplx
• know less that i did -- i know how much less i know
• PiRogue - jetstream - flamegraph
• (pi rho)
• ALU -- arithemic logic unit
• work on for next week -- too much to do. nothing related to here.
• want to try the bsky feed, firehose
• sophia made a guy from

https://github.com/mcint/dotfiles/blob/neon-mbp/utils/netcheck.bash

Questions, Discussion, or Coworking

• [Issue]

For next time

Readings & Exercises

• Readings
• Exercises

Join online

• Try it yourself!
  • Join libera.chat #nb-meetup-infra

https://www.noisebridge.net/wiki/Meetups/Infra