Post AwNVr2OG6m4cjU9arY by mycal@noauthority.social
(DIR) More posts by mycal@noauthority.social
(DIR) Post #AwNUbEeQegfiAgcOKO by picofarad@noauthority.social
2025-07-22T02:45:51Z
0 likes, 0 repeats
Linux servers can handle 3 million opened sessions per second.but somehow, we need k8s to "scale" serving text files and a limited array of compressed binary formats!
(DIR) Post #AwNVr2OG6m4cjU9arY by mycal@noauthority.social
2025-07-22T02:59:55Z
0 likes, 0 repeats
@picofarad define sessions, I'm calling bullshit on this. Best I know is meta's signaling servers, 200-500K concurrent connections, but there are concurrent low bandwidth mostly idle connections with little churn. 3M per second churn not happening on any one server.
(DIR) Post #AwNXUqKGTzMIqwpWxk by picofarad@noauthority.social
2025-07-22T03:18:19Z
0 likes, 0 repeats
@mycal https://www.phoenixframework.org/blog/the-road-to-2-million-websocket-connections ten years! TEN!Like, if you need to inspect all traffic on a saturated 10gbit link, as i've done 13 years ago, you tell the (probably linux based) core switch to mirror ports and feed the ports to a linux machine with at least 3 but ideally 4 10gbit NICslike, how can core switches and these traffic inspection servers handle all this?microservices rot the brain of everyone involved.
(DIR) Post #AwNZlbCYATcOmDL636 by picofarad@noauthority.social
2025-07-22T03:43:46Z
0 likes, 0 repeats
@mycal also, from 2011 60k new sessions per second on pretty junk hardware, with 2mm max open sessions. in ~2011. 14 years ago.Palo Alto Networks PAN-4050 also could handle 8x64k "buckets" of ports. using CENTOS. that's ~500k open NAT sessions. see second image, 2014 (11 years ago)PAN and F5 probably use fedora now (doubt rhel, but who knows), per the internet. Doesn't matter, it's still linux.
(DIR) Post #AwNb1rwjmJXNIWGUfQ by picofarad@noauthority.social
2025-07-22T03:57:55Z
0 likes, 0 repeats
@mycal pfsense claims their non-kernel networking can do 20 million packets per second, per xeon core.that's a lot more than 250k ;-)
(DIR) Post #AwNbLbx4xH5dU1q38K by picofarad@noauthority.social
2025-07-22T04:01:28Z
0 likes, 0 repeats
i know it's hilarious to lash out and say i flip burgers and merely speculate, but i personally built some of these traffic inspection servers (using gentoo), directly connected to core switches that were serving >500,000,000 MAU15 years ago
(DIR) Post #AwNdxDmDUC6rPxorKK by chromeratt@noauthority.social
2025-07-22T04:30:41Z
0 likes, 0 repeats
@picofarad that capacity is exactly why kis is valuable. It allows all that through put to be used by multiple applications early with reduced risk of configuration or version conflicts and scheduling issues.
(DIR) Post #AwNecak70HVKWDYteq by picofarad@noauthority.social
2025-07-22T04:38:09Z
0 likes, 0 repeats
@chromeratt Yes, broken organizations will produce broken code that needs "orchestration software" it's called Conway's law. If you have a team of 12 people servicing a half billion unique visitors a month your tech stack and architecture is gunna reflect that. We had ~300 average VMs handling that load, and duplicate of all the core switches, firewalls, traffic inspection, IDS, SSL termination, etchorizontal scaling is only necessary for *some* workloads; but people never name the right ones
(DIR) Post #AwNefC7P0qnX14hGoS by picofarad@noauthority.social
2025-07-22T04:38:38Z
0 likes, 0 repeats
https://en.wikipedia.org/wiki/Conway%27s_law
(DIR) Post #AwNhm0ETtzJm2hwiTg by professionalbigot69@poa.st
2025-07-22T05:13:29.218664Z
1 likes, 2 repeats
@picofarad @mycal >he's still using 10gig nicsWe live in the now, grandpa
(DIR) Post #AwNhqhMLV1W2xMrbdY by picofarad@noauthority.social
2025-07-22T05:14:19Z
1 likes, 0 repeats
@professionalbigot69 > 10gbit link, as i've done 13 years ago
(DIR) Post #AwNiPVty5gelYZBveK by professionalbigot69@poa.st
2025-07-22T05:20:37.571405Z
0 likes, 1 repeats
@picofarad You could prob boot off a disk a block away and still have decent loadsThat's sick
(DIR) Post #AwNisaizEqBQucv7Ls by picofarad@noauthority.social
2025-07-22T05:25:52Z
0 likes, 0 repeats
@professionalbigot69 unfortunately only 850nm sfp+ , the 10km stuff is more expensive, although i haven't looked into it in a long time.oh they're $40 now. not sure about compat, will look in to it. that is to say don't think i haven't thunk it, some people near me have much better HVAC than i do
(DIR) Post #AwOWwWvMJ0w9zYmLYm by me@pleroma.boingo00.wtf
2025-07-22T14:46:49.967194Z
0 likes, 0 repeats
@professionalbigot69 i dont need it, i dont need it…
(DIR) Post #AwPTmTFizofl0LmCqe by mycal@noauthority.social
2025-07-23T01:46:07Z
0 likes, 0 repeats
@picofarad Nothing I see on this thread says anything close to 3M per second on TCP connections on one server.
(DIR) Post #AwPUJoJlhH59tSZ7ce by picofarad@noauthority.social
2025-07-23T01:52:08Z
0 likes, 0 repeats
@mycal sorry you can't read, the first thing i linked was 2 million from 8 years ago, so
(DIR) Post #AwPUTyaCO2Ymcr68hM by picofarad@noauthority.social
2025-07-23T01:53:58Z
0 likes, 0 repeats
@mycal i'm literally on the horn with the network engineer that runs the network and the machines traffic inspection linux servers and he said "HAHA IS THAT GUY STILL ON ABOUT THAT?"
(DIR) Post #AwPaWaS79uZmNdwdfs by mycal@noauthority.social
2025-07-23T03:01:39Z
0 likes, 0 repeats
@picofarad 2M concurrent connections is nothing, but 2M per second is a whole other thing. Still bull shit.
(DIR) Post #AwPsC050TqagxJTAlE by picofarad@noauthority.social
2025-07-23T06:19:39Z
0 likes, 0 repeats
@mycal look, you're actually starting to annoy me:> This let us to reach 2M subscribers without timeouts and maintain 1s broadcasts.from the fucking article i keep telling you to readwhat does 1s broadcasts to 2 million subscribers from a single linux machine, TEN YEARS AGO, mean to you?ten years ago a single linux machine could publish to 2,000,000 subscribers *Every second*that's 10x more than whatever you spoke of yesterday.now, knock it off.