kliras posted:AV1 encoding is getting wider adoption on YouTube; this HWUB video has AV1 across all resolutions: If they're going to switch, I'd at least like to see statistical and perceptual similarity comparisons between the source, the old compression, and the new compression - although I doubt we're gonna get that. Combat Pretzel posted:I doubt ISPs have stopped their "Youtube/Netflix needs to pay for the bandwidth they use on our networks" shpiel. It's probably partly related to that. And IIRC back when Covid started, the EU mandated lower video bitrate to make room on the interwebs for home office data traffic. Not sure if that's still in effect. Klyith posted:At low resolutions the AV01 encodes get more bitrate than the VP9 & AVC encoders though. And AV1 is really good. I'd expect a reduction in bitrate by 10% versus VP9 still results in higher quality. I know a little less about video compression than I do about audio compression, but anyone with a confetti cannon or the equivalent After Effects filter can demonstrate how quickly things can get bad the same way that songs (even officially-uploaded ones) can end up sounding like vocal fry: https://www.youtube.com/watch?v=Gtyi1vvXxKY (Towards the end of the song, the crescendo is so powerful that it breaks most audio compression algorithms, and certainly all the ones offered by youtube - if you listen to a good FLAC rip of the song you'll hear very clearly what I'm talking about) BlankSystemDaemon fucked around with this message at 20:17 on Jan 18, 2022 |
|
# ¿ Jan 18, 2022 20:04 |
|
|
# ¿ May 14, 2024 07:07 |
Inept posted:Netflix probably has well under 1% of the data that Youtube has though. https://www.youtube.com/watch?v=_o-HcG8QxPc
|
|
# ¿ Jan 18, 2022 20:24 |
Combat Pretzel posted:Netflix also have cache servers in data centers of a lot of ISPs. Although I didn't mention they're free, which I wasn't aware of.
|
|
# ¿ Jan 18, 2022 20:25 |
repiv posted:yes netflix pushes a lot of data around, but it's largely pushing the same data to many different users, so that data can be cached at the edge You can probably guess how it went. The point was that they're still moving a shitload of traffic.
|
|
# ¿ Jan 18, 2022 20:30 |
Kerbtree posted:This video is not available? It's Comforting Sounds by Mew in case you want to look it up for yourself.
|
|
# ¿ Jan 18, 2022 23:11 |
Klyith posted:https://en.wikipedia.org/wiki/AV1#Quality_and_efficiency, there are results from a bunch of different research papers you can dig into Also, did you read any of the articles? Almost all of them measure structural similarity (or even more laughably, PSNR) whereas what I'm asking for is both that and perceptual similarity (preferably at the same time, since that tells more than either alone does). All of the research done on AV1 is on researcher-chosen videos (which can be an issue, but let's assume it isn't) as well as settings that've chosen based on whatever optimizes for the best quality at a given bitrate. That latter is the important part, because it's not what Google are targeting; even if we knew their exact encoding choices (while I haven't looked extensively, I've not found anything conclusive - but if you upload something following exactly their recommendations it still gets re-encoded, so they're doing something else), I'm pretty sure we can guess that they're more interested in what gets them the lowest bitrate while still having passable quality. That's a distinction that matters when you're measuring PSIM. As for Comforting Sounds, the mastering of the song isn't the problem as it sounds great on the Frengers album or ripped as FLAC - it just happens to be a song where the crescendo contains so many things that the psycho-acoustic properties of most lossy codecs break down unless you do CBR at the highest bitrate. However, it's also a good test for this exact reason, because you don't need a blind ABX test to hear the differences in the FLAC encoding vs VBR in the low-140kbps- ie. what the OPUS encoding is for the version I linked, but even if it was AAC-LC (which is what youtube recommends), it still sounds like vocal fry. BlankSystemDaemon fucked around with this message at 17:02 on Jan 19, 2022 |
|
# ¿ Jan 19, 2022 16:52 |
|
|
# ¿ Jan 19, 2022 20:43 |
I tried elucidating some of the more glaring issues over in the digital packrats thread.
|
|
# ¿ Jan 29, 2022 20:09 |
Former Human posted:Remember the tape drive Linus hyped in 2018 for backing up data? I guess they never actually used it. Especially because an actual tape library with a robot is the one way to handle data in these orders of magnitude, not to offload it to the
|
|
# ¿ Jan 30, 2022 11:32 |
Tape and tape libraries are one of those things that are like fax - if you're paying someone for eFax services when fax is supported by every modern OS for at least a decade, there's probably other people who'd like to take you to the cleaners. BareOS is better than whatever enterprise tape backup software you can buy, and it's free.
|
|
# ¿ Jan 30, 2022 22:28 |
Charles posted:Or maybe it's just entertainment and we have fun learning things along the way! BlankSystemDaemon fucked around with this message at 08:10 on Feb 1, 2022 |
|
# ¿ Feb 1, 2022 08:06 |
Former Human posted:A Linus "we lost all our data" video must naturally be followed up with a "look at our NEW server (which is basically the same as the last one that failed)" video. (Source)
|
|
# ¿ Feb 4, 2022 11:19 |
Klyith posted:%1-%2 AFR is the overall average for their drives. I'm not sure why BSD found that model to be lol-worthy, because it seems perfectly normal. The point being, that Seagate are still doing worse than their competitors, and these are the drives that the youtuber was saying there was absolutely nothing wrong with, because he'd received the drives from Samsung so they were perfectly fine and definitely not worse than the competition, nope, not at all. That is what I was laughing about. BlankSystemDaemon fucked around with this message at 21:40 on Feb 5, 2022 |
|
# ¿ Feb 5, 2022 21:36 |
Klyith posted:Their AFR for their entire pool has ranged between 0.93% and 1.89% over the past 4 years. That particular model has been above-average every year it's been on the charts. It's not brilliant, and Seagate seems unquestionably worse than HGST for reliability -- hard to say anything about WD since Backblaze almost never uses them. The pool failed because one of those raidz3 vdevs had more than 3 drives fail, not because all the drives failed. Their storage was designed to fail, even before we begin talking about the drives they were given and them insisting that the drives are fine despite the fact that there's evidence (higher than average/mean AFR) to suggest that those drives aren't fine. Also, please re-read what BackBlaze mean by Annualized Failure Rate - it's not the same as Annual Failure Rate, and accounts for the power-on time already, so you're not supposed to factor it again. And speaking specifically of the higher AFR for the Seagates, the article I linked explicitly talked about how the AFR compared to the Q3 numbers dropped after they updated the firmware - something the youtuber is unlikely to have done given that they didn't update the rest of the system either). BlankSystemDaemon fucked around with this message at 13:10 on Feb 6, 2022 |
|
# ¿ Feb 6, 2022 12:48 |
If he wants something that's bound to fail, he should be using BTRFS.
|
|
# ¿ Feb 6, 2022 16:29 |
Klyith posted:Ah, and that's consistent with their rebuild attempt throwing errors on every drive in the array? I don't have knowledge in this subject so I was taking Linus's word that errors distributed across all drives meant it wasn't a single drive failure. If a device experience too high of a rate of change for either of those columns, or if the count is too high, the device is marked as as either degraded or faulted. If the device disappears completely (presumably because it catastrophically failed), or if it's taken out of rotation with zpool offline, it's marked as unavailable or offline respectively. All states except online mean that the fault tolerance of the entire pool is compromised in some fashion. If memory serves, their hardware RAID was a RAID5 of RAID0s (ie. RAID05 instead of RAID50, which is fairly normal although RAID60 moreso) - which itself is pretty indicative of their competence. 2% AFR is still double the AFR of the mean for their entire fleet of disks, and some of the disks have an order of magnitude lower AFR at between 0.10% and 0.20%. While it's not the 6-10% AFR that we've seen for some manufacturers, 2% AFR is still bad.
|
|
# ¿ Feb 6, 2022 19:47 |
Combat Pretzel posted:He just published a video about using Linux' filesystem abstraction layer, that allows for just about anything to act as block storage, for swap space to enable "10TB of RAM" or some poo poo. gently caress me, they're running out of topics or something?
|
|
# ¿ Feb 6, 2022 19:49 |
Charles posted:You could just unsubscribe.
|
|
# ¿ Feb 6, 2022 22:01 |
Dr. Video Games 0031 posted:entertainment youtube channel: "here's a silly thing you can technically do but probably shouldn't" Dr. Video Games 0031 posted:...Did you see the video? The part where linus dresses up as a mad scientist and cackles maniacally might clue people into it not being serious tech news. They then go on to talk about why it's a dumb idea in a manner that's kinda interesting. They aren't telling anyone to do this. Even if they dress up like mad scientists, I can absolutely see how someone, who doesn't understand the catch-22 of swap (ie. the act of paging something to swap, on a filesystem or anything else that isn't a disk/partition, requires allocating more memory which in turn requires paging more to swap), might well think that that's actually a good idea. This is further complicated by their genuine incompetence when it comes to anything more serious than using a screwdriver - like their business storage solutions always breaking, their attempts to do multi-tenancy virtualization solutions, and so on and so forth. BlankSystemDaemon fucked around with this message at 10:53 on Feb 7, 2022 |
|
# ¿ Feb 7, 2022 10:51 |
forest spirit posted:about linus being the best That must suck.
|
|
# ¿ Feb 7, 2022 19:01 |
It's worth noting that they're working with uncompressible data, since video is already compressed - so inline compression features from tape or a modern filesystem like ZFS isn't going to get anything - especially because lz4 and zstd features early-abort mechanisms, where it stops trying to compress records, if it can't compress them at least 14%. Since it's a data archive, I think nearline storage (either in the form of a tape library with robotic arm, or a MAID; a massive array of idle disks where spinning rust is spun-down) is the way to go. That is, unless they also have some kind of IOPS requirements, in which case what SlowBloke mentioned is more the neighbourhood they're looking at. It's also not clear to me if it's 1.2PB allocated space, or if that 1.2PB is just what they think they're getting if they're assuming that "14TB" drives are 14TiB and aren't accounting for filesystem metadata, distributed parity, slop space, and other things that's gonna affect the allocated amount of space.
|
|
# ¿ Feb 8, 2022 12:16 |
Getting data out of the butt is exorbitantly expensive, which is entirely by design If I was building for max efficiency, I'd look into getting a 36-bay front-loading chassis from Supermicro, and 9 of these to fill up the rest of the 42U rack for a total of 846 disks per rack. The advantage they have is that you're not moving 60-90 drives at a time, which is the case with all the top-loading drives, as these disk shelves allow for easy servicing of two disks at a time without moving any other drives at all. Assuming 16TB disks, with 9 devices per raidz3 vdev plus a bunch of spares, that's over 7PiB of usable storage. EDIT: And if you go with DRAID, you can probably get more than 10PiB per rack. BlankSystemDaemon fucked around with this message at 14:19 on Feb 8, 2022 |
|
# ¿ Feb 8, 2022 14:02 |
Arivia posted:You really obviously didn't watch the video, the 1.2PB is just the theoretical maximum of the drives. When they get things set up in their NAS installation it's in the 900 terabyte range and there's a couple of jokes about doing a bad setup to make the magic 1 petabyte number appear but Linus says not to bother, to go with the right thing. They know they're not getting an actual petabyte of storage out of that thing.
|
|
# ¿ Feb 8, 2022 14:15 |
Arivia posted:My apologies, comment retracted then.
|
|
# ¿ Feb 8, 2022 14:21 |
If my napkin calculations are correct, they could've bought two 90-bay SAS disk shelves and a 36-bay chassis, and gotten 1.17PB usable space with 10TB disks - and they'd still have 9 disks left over for spares. And quote is not edit.
|
|
# ¿ Feb 8, 2022 14:26 |
Do 100 racks and get 1EiB, or 100000 racks and get 1ZiB!
|
|
# ¿ Feb 8, 2022 20:14 |
The Grumbles posted:https://www.youtube.com/watch?v=_YGA8ufRjiY
|
|
# ¿ Feb 9, 2022 13:37 |
Phone posted:thanks algorithm
|
|
# ¿ Feb 11, 2022 18:36 |
Grapplejack posted:It's kind of shocking how many people are leaping at the 'one bad apple' defense for newegg, despite the fact that this went through so many hands and departments that it's impossible for that to be the case, let alone the company culture that promotes this sort of poo poo in the first place. This isn't even talking about newegg's shuffle thing where they are bundling actual trash with GPUs to get sales on otherwise unsellable items
|
|
# ¿ Feb 12, 2022 19:54 |
Mr.Radar posted:It's because they're a media company that gets ~75% of their revenue from ads and it would be "hypocritical" for them to support ad blocking (even though their employees would be a freaking molecule of water in a bucket). Jeff Fatwood posted:Linus outed himself as a complete shitter to me on a previous WAN show when he said he'd be personally offended if his employees unionized because he treats them so well. Good treatment and worker rights are OK to him as long as he gets to arbitrate and feel good about it.
|
|
# ¿ Feb 19, 2022 12:25 |
Former Human posted:I've always suspected that the sweepstakes on playr.gg (I guess they're Surf Giveaways now) worked the same way. Only the entrants with the highest followers and subs on Twitch, Twitter, and Instagram would win prizes. It didn't matter how many entries you had in a given contest, if you were a low level streamer or influencer you weren't getting dick.
|
|
# ¿ Mar 2, 2022 06:53 |
VostokProgram posted:this has to be illegal though right? gambling is extremely regulated, and i thought giveaways are the same way And what happens in situations where the infrastructure the company is using is hosted and/or the people who own the company are outside the jurisdictions of the country in question?
|
|
# ¿ Mar 4, 2022 17:07 |
Warmachine posted:11:24 the MC sales guy asking Steve if there's a meaningful difference between CL16 3200 and CL16 3600 There's almost no real-world use-case where you see peak bandwidth, especially not if you're doing memory-intensive workloads - so what matters a lot more is the sustained bandwidth which depends on too many factors to easily list. If you wanna know more on the subject, I'd recommend the paper a performance & power comparison of modern high-speed dram architectures by Shang Li, et al. BlankSystemDaemon fucked around with this message at 22:59 on Mar 7, 2022 |
|
# ¿ Mar 7, 2022 22:56 |
If an unnamed benchmark is showing 10% difference with 5% margin of error and a significant standard deviation at a confidence of 10-50%, that's about as useful as just pulling any number out of your rear end. If you begin factoring all of those and ensure that you have enough data points for it to be statistically valid, along with things like rebooting between each test, a lot of that measurable difference suddenly starts to disappear. BlankSystemDaemon fucked around with this message at 17:46 on Mar 8, 2022 |
|
# ¿ Mar 8, 2022 17:42 |
Linus being wrong is the practically so common, that the word Linus might as well translate to wrong. Heck, he even pronounces his name wrong compared to how everyone in the Fenno-Scandic region, where it comes from, pronounces it.
|
|
# ¿ Mar 12, 2022 09:36 |
Former Human posted:How does Linus Torvalds pronounce it??? There's also some video floating around where he pronounces it the Americanized way, probably because the Linux Foundation is incorporated as a 501c6 in the US (ie. a board/chamber of commerce, where corporations pay to enter), and as such has a trademark. BlankSystemDaemon fucked around with this message at 15:57 on Mar 12, 2022 |
|
# ¿ Mar 12, 2022 15:52 |
Yeah, that's the whole trademark deal I was talking about.
|
|
# ¿ Mar 13, 2022 13:09 |
Rinkles posted:https://www.youtube.com/watch?v=dtG9I3mZlJo I'm okay with sponsored content, so long as it's clearly labeled - which I trust Ian to do; partially because he's legally obliged, of course, but also because he's managed to retain his integrity over a very long period of time. EDIT: Also, I just noticed. The logo for his company is a rainbow - I wonder if there's anything to that beyond the iridescent surface of wafers. BlankSystemDaemon fucked around with this message at 21:55 on Mar 14, 2022 |
|
# ¿ Mar 14, 2022 21:51 |
VostokProgram posted:Would be extremely funny if he got a job at GN
|
|
# ¿ Mar 14, 2022 23:47 |
|
|
# ¿ May 14, 2024 07:07 |
Adolf Glitter posted:Watched the whole video now. I totally see him as an analyst, the analyst-influencer thing puts my teeth on edge, but maybe that's just me being old.
|
|
# ¿ Mar 15, 2022 00:24 |