ServerMeta

Be sure not to miss my interview on #nodejs , #rust and #wasm! sprkl.dev/performance-ru…

@axboe I know what you mean 😜 I had the same experience while writing a tutorial about io_urint and nvme 😂

Decided to write a quick doc on "io_uring and networking", and 2h later I'm at 7 pages and have to stop myself. Coming soon...

@Franc0Fernand0 Nice, should also mention rendezvous hashing which is the current frontier of research

@elonmusk ok we are starting with the blaming others strategy. Really a good manager.

Twitter experienced slight degradation of service today from an old 3rd party tool used to block accounts that had no rate limit (sigh). Should be fixed now.

@ThePrimeagen This way you have batching, much less memory overhead, and you don't need the WASM bindings. Also the wasm worker can maintain its own vdom, to minimize changes.

@ThePrimeagen Wasm will compute the desired DOM changes, and then store them in the array buffer and signal the main thread. The main thread will apply the changes and save the operations result in the main buffer, and signal wasm once done. 2/3

fellas, its looking bad for wasm perf against react on the client side. Fs in chat (still rendering at +200 fps which is enough (with 1920 updates per frame))

@GavinRayDev @jorandirkgreef @axboe is there any benchmark? I'm always skeptical about windows performance

@jorandirkgreef @axboe This is so awesome!! On the Windows bit -- there's a very recent thing they've done called Io_Ring, which has comparable performance to io_uring and should be preferred to io_uring if the support is present on the target host: windows-internals.com/ioring-vs-io_u…

Writing your own event loop over io_uring is such an awesome way into systems, with io_uring’s unified first-class API for async disk and networking. Shoutout to @axboe for making single-threaded control planes cool again 😎

@FiloSottile what do you think of rustls? Is it safe?

@_ericelliott which language was the code in?

Doing a code review right now - saw some code I did not immediately follow, so instead of mentally running the code line-by-line like I normally would, I pasted into GPT-3 and asked questions. GPT-3 explained the code to me - correctly. 🤯

@matteocollina @rakyll on the other hand gcp doesn't have an equivalent of dynamo

@matteocollina @rakyll becuase AWS is a huge company, and very slow. When they release software they need to support millions of customers. Google on the other hand has been working on it for several years (from the launch of kubernetes).

Cloud Run just works and its only downside is it's not an AWS product.

@sc13ts with io_uring I use kprobes

What do y'all like for tracing asynchronous systems on a single machine? I want to see a global timeline, draw a waterfall chart, and also get access to the raw data to write queries. Lots of options for big distsys, but can I get away with something simpler on one machine?

@phil_eaton I would say bitcode allows for easier implementation of big int, big decimal and makes it easier to support exotic architectures, like GPU/TPUs have usually bigger word size.

I'm the dummy wondering: why on earth is it LLVM *bitcode* (not bytecode)? stackoverflow.com/a/36106255/150… Is this it? That you can pack more information in? But most computers that run the LLVM process are byte oriented so what's the point? Or what am I missing? 😀

@rhein_wein is 40k is reasonable amount? Seems a bit much to me. The real cost maybe is the salary of the new employee, which will most probably be higher than the fired one.

Employee A is a solid performer and asks for a 10k salary increase. But they're already at the top of the salary band, so it's not possible. So they find a new job that pays them 15k more, and the company ends up spending 40k to hire and train a replacement.

actually one more question: - can you use fixed buffers with NVMe commands? (maybe you would prefer these questions as a github issue?)

- Can I use fixed buffers with recv? Or is it better to use read with sockets? - Does io_uring do any batching behind the scene? Maybe with SQPOLL can I expect requests to be batched? - What do you think it's more important, fixed buffers or batching? 2/2

Hey @axboe it seems my example is becoming very popular, and many people contacted me for questions and more examples. I will try to collect tips and best practices in a tutorial but I see there's a lot of stuff to investigate :) Some questions to start: 1/2

@samlambert distributed transactions, modern sharding with rendezvous

What database topics would you like to learn more about? We have some in depth blog posts planned and I want to add to the list.

After 4 months of hard work, I made an NVMe io_uring example work. So hard, but so happy that I made it. I feel like I improved a lot. Thanks to @axboe for building io_uring and Ankit Kumar for helping me. github.com/espoal/uring_e… #io_uring

I partecipated in Lars Wirzenius Rust training course, it was super interesting to see a seasoned engineer way to use Rust. liw.fi