Ankush Jain

@ajay9470 I moved to YT Music India, some family plan or something via dad's account I don't even remember but it was cheap. Anyway my daily jam these days is lofi music to plan a thesis proposal to.

You know it's been too long when even Spotify tells you to graduate

Looking to recruit "one" PhD student (in either ECE + CS @ UIUC) with STRONG interests in ML+Systems+Security, or demystifying foundation models. Deadline: Dec 15. Ping with any questions. Please RT and help amplify?

@axboe @rflaherty71 Would it be accurate to map this to be in the ballpark of the CPU overhead of ~200K IOPS (100G/64k) in the NVMe world + some DMA setup/teardown cost?

@rflaherty71 Yeah they are large, kperf uses 64k by default. But as mentioned in the post, it's just a single cpu core, not even user the HT sibling.

100G is boring, so I went 200G instead on the rx side. Implemented zc send in the test app (kperf) as well (io_uring already supports this). Single CPU thread, maxes out 200G link as well. See below for basic details: @kernel.dk/" target="_blank" rel="nofollow noopener">lore.kernel.org/io-uring/9bbab…

@GuilhermeAmadio @axboe It depends on your profiling requirements - if they're simple something like ifstat would suffice. If they're complex something like tcptop from bcc-tools (or anything in the BPF world) can be extended to get almost anything you'd want.

@axboe I see, I was just wondering what you used to get the transfer rate, and if you look at off-CPU time, scheduling, etc. Today I wrote a simple shell script to print transfer rate over time, for example, because dstat only does 1s intervals, and I wanted a bit higher frequency.

David recently posted the first non-rfc version of the io_uring zerocopy receive work, you can find it here: @LQ3V64L9R2/T/" target="_blank" rel="nofollow noopener">lore.kernel.org/io-uring/ZwW7_… I noticed Broadcom has public fw that supports this now, so really very few (and minor) dependencies left for this work. I ran some testing. ⬇️

@RoshalSophia I use Zotero + bib auto-exports + the Zotero connector browser plugin - all bib items are automatically accumulated in the appropriate bib files as per your rules.

Why do certain (unnamed) websites force you to download the bibtex citations rather than just being able to copy? No one needs individual files for each citation and my already poor old laptop that constantly tells me I'm out of space really doesn't need more junk files on it :/

@AbhirathB It was just me saying that there are low-hanging fruits (alpha/productivity gain from tools) everywhere, and all will eventually adopt them but: 1) Incomes need to rise above a certain threshold 2) Most folks will still be slow to adopt. One person will, then a slow cascade

@schwifty50 Like in this case what is the alpha and incentives? I didn't quite figure out the mapping.

India's offering to the Global market is nearly entirely based on Labor Arbitrage. So squeezing every last bit out of Labor is the critical piece of Indian business. The central cultural challenge isn't bosses being "exploitative" but that the culture isn't interested in ...

@AbhirathB I would encourage you to play around with this website: oec.world/en/profile/cou… Compare India's export profile and economic complexity to other countries in its PCI bracket. We've done great with 90th ptile industries, but not really 60-80th ptiles.

@AbhirathB But labor arbitrage usually refers to sweatshop-type labor. With BPOs/call center-style jobs, it is not just that it's cheaper, but also that folks can speak English and all. There's an element of skill. Until the recent experiments with PLI in mfg, we haven't been great at that

@AbhirathB In general I am skeptical of this trend of prematurely identifying cultural root causes as the driving forces for certain behaviors. The gap between what is possible and what exists is alpha. Alpha creates incentives, which all humans respond to, but with varying amts of interia

@AbhirathB I would disagree. India does not really do a great job at labor arbitrage (especially compared with peer countries, like BD). Successful Indian exports like pharma/IT/autos are not low-tech. These can be explained by inertia to ideas. (Hacker culture being the extreme opposite).

@ffledgling @Android @Apple It's inevitable, but conventional automakers don't have the rapid iterative loops of tech. Tesla is able to do things faster because there's compute onboard and software-defined things and processes that deliver rapid OTA updates.

I wonder if @Android Auto or @Apple Carplay will eventually start supporting netflix/youtube natively. As EVs start to proliferate and more people start spending time at chargers just waiting, it'll probably become more of a sore-point. (Tesla owners, please don't gloat)

@mnwsth I think economic incentives guard against that. Your customers will pay per core, and it is in your interest to minimize the overhead.

“The current cumulative capacity of all data centres pan India is roughly 950 megawatts (MW) and there is an opportunity to double this in the next two-three years.” It still feels a little weird to me that DC “capacity” is measured in Megawatts. What if there are a lot of badly designed DCs with high PUEs? Then the power consumption would be inversely proportional to compute capacity. m.economictimes.com/tech/technolog…

@ajay9470 You're not?? Did something specific happen between 9 and 11?

@schwifty50 Nice! I am not surprised

GCC 11 feels SO MUCH FASTER than GCC 9 this is ridiculous.

@ajay9470 Compile time for decent-sized C/Cpp codebases, on somewhat ancient Sandy Bridge Xeons. Probably went down by half if not more.

@schwifty50 compile time or run time?

@ajay9470 Check auto shop loaner tool sections? Pittsburgh has a tool library, not sure if you have something like that.

@schwifty50 Also right now it's working fine, just making slight noise and there is a slight play in the bearing. I was hoping it would just be gunk/dust but after taking it apart it's definitely bearing damage

It's Scooter Repair Saturday. Bearings on the rear wheel is fucked, someone help ☹️

@ajay9470 Ok nvm for scooters looks like you can just hammer them out/in. I'd get a caliper, measure ID/OD/Width, look up the size ####, and order replacements by that number.

@ajay9470 Bearing sizes are very standard AFAIK and replacements are cheap, may need some extractor/press tool thing to get them out and put new ones in. You can get a bunch of loaner tools from AutoZone for free.

@ajay9470 I had #2 in my header files and was getting "undefined reference" linker errors and couldn't figure out why it wasn't generating the code, turns out I had to remove the <>

@ajay9470 No my confusion was, for: template<typename T> func(T), between: 1. template func(int); 2. template<> func(int); #1 is an instantiation of the template fn body for ints. #2 is saying that you'll provide a different body for ints. I thought they were the same.

This should be obvious and you shouldn't be touching a computer if you don't get it and it is okay for the compilation to just throw a linker error when you mix them up despite clangd autofilling one where you needed the other