Oh, lovely, you're a cheery one aren't you?
Friday, May 22
Potato Ratio Edition
Tech News
- Twitter has released its new Brian Stelter Is Being Ratioed Again button, which protects Brian Stelter from being ratioed again.
Think I'm joking?
It doesn't show up on the embed, but you can't reply to that.
- There are no Passmark results for the 10900K just yet, but while I was looking I discovered that there are results for the Ryzen 3900. (CPUBenchmark)
The 3900 is a 12-core 65W OEM-only part. There's no consumer equivalent, because you can just buy a 3900X and set it to 65W mode in your BIOS, and still retain the option to put it back in full 105W mode at a later date.
I thought this was worth comment because the 3900 delivers 95% of the multi-threaded performance of the 3900X. I'm sure most users don't begrudge that extra 40W. But the 65W performance numbers indicate that AMD could release a 24-core desktop part on 5nm and keep it within their 105W nominal TDP and running at current clock speeds. Plus two generations of IPC improvements, of course.
- The public does not understand logarithmic graphs. (LSE)
No shit.
- Hot Chips 2020 is coming. (AnandTech)
Online-only this year, of course; big conferences will be the last thing to go back to normal.
Highlights this year are IBM's Power10 and z15 processors, Intel's Ice Lake server parts, Marvell's 96-core 384-thread Arm server processor, and a 4096-core RISC-V processor.
- The ASRock Z490 Taichi has two PCIe 4.0 slots and a PCIe 4.0 M.2 slot. (PC Perspective)
The ASRock Z490 Taichi is an Intel socket 1200 board.
Intel doesn't have any PCIe 4.0 parts.
That pretty much confirms that Rocket Lake - Intel's 11th generation parts, based on the Ice Lake core - will be PCIe 4.0.
- The Alienware M17 R3 has everything you might need in a software development laptop. (Tom's Hardware)
Big bright 4k display? Check.
8 core CPU? Check.
2TB of NVMe SSD? Check.
32GB RAM? Check.
RTX 2080 Super? Check.
2.5 hour battery life? Oops.
It has PgUp/PgDn keys - and a numpad - so it has that covered. It also was a 330W charger.
- Eight years after the developers announced they were winding down work on Terraria, the game has received yet another huge update. (Terraria)
I'd like to get back into it; I stopped playing when a freak parallel-world respawning incident blew up my fish statue.
- Intel's 10900K hits 7.7GHz on all cores. (Tom's Hardware)
All it took was a bit of liquid cooling.
Liquid helium.
Liquid nitrogen is cheap and readily available and fairly easy to handle. Liquid helium is none of that.
Still, hats off to the crazy bastards.
Posted by: Pixy Misa at
02:43 AM
| Comments (5)
| Add Comment
| Trackbacks (Suck)
Post contains 477 words, total size 5 kb.
Thursday, May 21
Ten Core Toaster Oven Edition
Tech News
- Intel's Comet Lake is here. (AnandTech)
The i9-10900K is usefully faster than the i9-9900K. Largely due to having ten cores rather than eight. It's specced with a 5.3GHz top boost clock but a 280mm radiator is not enough to ever actually reach that.
Despite not hitting the rated boost clock, it did hit 254W sustained under load, compared to 144W for a 3950X. But for multi-threaded workloads the 3950X runs rings around the 10900K.
Other than that, well, it's Skylake on 14nm. It's not actually bad unless you care about power consumption in which case it is very bad indeed.
- If you're more interested in Linux, the 10600K and 10900K get a workout there too. (Phoronix)
The 10900K does compete fairly well with the 3900X. Only problem:But at least according to the RAPL framework reporting based upon the Intel CPU metrics recorded, the actual peak power consumption of the CPU went as high as 380 Watts in extreme cases.
That's at stock speeds, not overclocked.
- Engineering samples of Zen 3 are already showing up. (Tom's Hardware)
Not early samples either - these are hitting 4.6GHz boost clocks. Release is rumoured for August / September which doesn't give Intel much time before they get clobbered yet again.
- Windows terminal has hit 1.0. (Microsoft)
With critical features such as... Animated GIF backgrounds and CRT simulation.
Um.
And tabs.
- ZFS vs. RAID. (Ars Technica)
ZFS wins. It can needs some tuning on spinning rust, but it allows that tuning - live, in production, on arrays already in use.
Posted by: Pixy Misa at
01:32 AM
| Comments (4)
| Add Comment
| Trackbacks (Suck)
Post contains 286 words, total size 3 kb.
Tuesday, May 19
That'll Slow The Fish Down Edition
Tech News
- A look at the ASRack X470D4U Ryzen server motherboard. (Serve the Home)
Coming soon to a mee.nu near you, because this is the exact board in Akane III.
It is an X470 motherboard, so it's PCIe 3.0, but for now that's not a huge problem. My SSD (a 3.2TB Samsung enterprise NVMe model) is also PCIe 3.0, and I'm not running a graphics card.
It has two M.2 slots, in an interesting and sensible configuration given the limitations of the chipset: One is PCIe 3.0 x2, and the other is PCIe 2.0 x 4. That means both have 2GB/s bandwidth, while leaving PCIe lanes available for the extra network controllers and BMC chip that are included.
Which does mean I don't get quite the full speed of the SSD (3.1GB/s read, 2GB/s write) but if I add another it should work exactly the same.
X570 would be much better, allowing two PCIe 4.0 x 4 M.2 slots. But even our new Threadripper-based production systems rarely break 1GB/s of I/O, and they run RAID-0 arrays of PCIe 3.0 x 8 SSDs.
Also, Akane's CPU is running at 33°C. I think it needs more work to do.
- Trillions of QNAP storage arrays are vulnerable to remote attack. (ZDNet)
Well, half a million anyway. Well, if they're connected to the internet. Or sitting on a corporate network with bad people wandering about.
The bug is in the photo app, which, um, runs as root. That seems ill-advised.
- Minecraft hits 200 million sales. (Thurrott.com)
That's a lot. It certainly justifies the price Microsoft paid Notch.
- Associated Press not only wants you to hand over your video of news events without payment, they want you to personally indemnify them against an resulting... Anything. (TechDirt)
And indemnify anyone they relicence or sublicence your content to, an infinite cascade of legal liability with no compensation.
Also, they're blocking anyone on Twitter who dares to criticise them over it.
- Samsung has a new 50MP camera sensor. (AnandTech)
Yes, that's half the number of their previous sensor, but those were pushing the limits of the wavelength of visible light, and using fewer but larger pixels is likely to work better overall.
Also, even with the previous 108MP sensor, Samsung phones defaulted to taking 12MP photos.
- Formatting code is hard. (Stuff With Stuff)
Mostly I just hit Ctrl-Alt-L.
Video of the Day
Yes. No. No? Yes. Yes? No? Yes! Mostly.
Looks like Ryzen 4000 is back on the cards for 400-series motherboards.
It's still not possible to support all AM4 CPUs on a single BIOS, so if you were planning to jump from a 1600 to a 4600 you may have a tricky upgrade process ahead of you. Which is why AMD didn't want to do it in the first place.
Posted by: Pixy Misa at
10:55 PM
| No Comments
| Add Comment
| Trackbacks (Suck)
Post contains 491 words, total size 4 kb.
Monday, May 18
Castle Is Scooby Doo Edition
Tech News
- The solution to deploying ZFS on closed VPS platforms is to use files. Just create an empty file of the size you want, put it somewhere it won't get accidentally deleted, and create your zpool using the file as you would normally use a block device.
Yes, it's an extra layer of indirection and it will be less robust than native ZFS filesystems, but you do get snapshots and block compression.
To try it out I created a 200GB file, write-protected the parent directory to keep it safe, created a zpool, and rebooted. Still there and working.
This is as simple as:
dd if=/dev/zero of=/fs/midori bs=1000000 count=200000
zpool create midori /fs/midori
You'll probably want to add:
zfs set compression=lz4 midori
To enable block compression. With a fast SSD and a decent CPU there is no really downside to this, and I typically see compression ratios of 1.7x with a mix of compressible and non-compressible data. Easily-compressed stuff like log files and JSON blobs can be 6x or more.
- The Great Bat Plague shows why we need Universal Basic Internet say idiots. (Tech Crunch)
Australia has Universal Basic Internet. The moment the legislation passed all progress on existing internet deployments stopped, and it took twelve years and two cancellations for the new network to reach me in suburban Sydney.
- MediaTek is well-known for its cheap mid-range ARM CPUs. The Dimensity 820 kind of pushes the boundaries on that. (AnandTech)
Four A76 cores at 2.6GHz and four A55 cores at 2.0GHz, built on TSMC's 7nm process.
- Google banned a podcast app from their Play Store for... Playing unapproved podcasts. (Reclaim the Net)
I wanted to check that. If you Google "podcast addict" the first result is a Play Store link... That doesn't work anymore.
And here's the developer:
- I've referred to Docker before as the world's least efficient package manager. Before I was forced to work with NPM. But it doesn't have to be as inefficient as it usually is. (Cloudreach)
A 2MB Docker image is something I can deal with.
- One of the X-37B space planes is off on an adventure. (BBC)
News articles generally refer to it in the singular, but this one mentions that there are (at least) two. We knew that not because the military announced it, but because it landed twice in a row without an intervening launch.
Posted by: Pixy Misa at
10:40 PM
| Comments (4)
| Add Comment
| Trackbacks (Suck)
Post contains 411 words, total size 4 kb.
I created a 50GB file, which ran at 288MB/s and was effectively CPU-bound, not I/O-bound.
I then used that file as a block device to create a ZFS pool - which you can totally do.
And it worked just fine.
I then deleted the file just to find out what happened, and that also worked. So, um, don't do that. If you're logged is as root you can trash the partition table and write directly to the raw block device, so don't do that either.
Performance during this experiment was just fine. Given the price, performance was awesome.
Posted by: Pixy Misa at
02:42 PM
| No Comments
| Add Comment
| Trackbacks (Suck)
Post contains 156 words, total size 1 kb.
No Infinitive Split Before Its Time Edition
Tech News
- I got one of those SSDNodes, um, nodes. 6 cores on a dual E5-2690 v3 host, 24GB RAM, 240GB SSD, for what works out, with a combination of three different deals, to $10.33 per month. That would get you one core, 2GB RAM, and 50GB of SSD at Digital Ocean.
On the one hand, Digital Ocean has a ton of features SSDNodes doesn't - the ability to resize your VPS on demand, private networking, block storage, firewalls, floating IPs, load balancers - and hourly billing. You only need to pay for what you need to pay for, from one hour to the next.
On the other hand, if you are prepared to commit a year in advance SSDNodes can be one-tenth the price for equivalent capacity.
Couple of limitations: There's no easy way to configure your VPS with native ZFS - though it is possible to create a ZFS pool on top of EXT4, which gives you all the ZFS filesystem features at the cost of CPU overhead. That's what you get if you follow their helpful LXD configuration guide.
And they require you to run one of their prebuilt Linux installs - no custom images or interactive installs. But as noted you can run LXD and configure your containers any way you want.
I/O performance isn't bad. I created a 12GB swap file and watched while it was written. Average throughput was 236GB/s; peak was 580MB/s. Reads measured byhdparm
run at around 100MB/s plus or minus 15%.
That's not amazing, but having 24GB of RAM available will go a long way to ameliorate that. And if you need disk performance they have NVMe nodes available. I didn't bother since I already have a dedicated server for that.
Unless I run into something unexpected - something far beyond just the CPU being shared with other users - it's a bargain.
- I've started working on retiring our backup server, which costs nearly as much per month as Midori - the SSDNodes VPS - does per year.
It contains 90 million files, dating back to 1994. No, I don't know how it has files dating back to 1994, but it does.
I remember now that when I got this huge 48TB ZFS server I copied all the old backups onto it and then uncompressed them. The idea was that I could create a global directory and remove all the duplicates; there's at least a dozen copies of everything in there.
I never actually got around to any of that, so now I need to recompress about 15TB of massively duplicated data.
This might take a while. But at least I can mostly just fire it off and leave it to run.
I've started with a relatively modest directory - just 125GB and 920,000 files. About 1% of the total. We'll see how that goes.
And while it runs I'm downloading all my GOG goodies.
- Intel has joined TSMC in plans to build a fab in the US unless it hasn't. (WCCFTech)
It seems that Intel is in talks to offer a dedicated foundry facility. They do offer foundry services, but almost all of their capacity is used for their own chips. Their one major customer was Altera, but then they bought them.
- GIFs considered harmful. (Danny Guo)
MJPEG is a thing, you know.
- Comparing Renoir to Ice Lake. (Phoronix)
How does the Ryzen 4700U compare with the i7-1065G7?
Considering that the Ryzen has eight cores and the i7 only four, you'd expect it to be a benchmark massacre. And so it is. Intel ekes out a few wins on benchmarks that don't multi-thread well, but AMD wins 88% of the comparisons by an average of 72%.
- What's new in WSL2? (Bleeping Computer)
If SQL is pronounced Sequel, isn't WSL pronounced Weasel?
Anyway: Real Linux kernel (which will fix some bugs I've noted when trying to use LMDB and compressed executables), much much faster I/O, and a long list of bug fixes.
- Someone is hacking supercomputers... To mine cryptocurrency. (ZDNet)
If this were the 80s, we'd be up to WWXVII by now.
- The latest version of Doom comes pre-doomed. (Ars Technica)
It includes the Denuvo kernel-level malware.
Posted by: Pixy Misa at
01:29 AM
| Comments (3)
| Add Comment
| Trackbacks (Suck)
Post contains 719 words, total size 6 kb.
Saturday, May 16
Dimetrodons Are Not Dinosaurs Edition
Tech News
- My Steam library is 4.3TB. Less than I was expecting.
Now to move on to backing up our backup server onto my new (old) Synology farm here at PixyLab. The server contains a historical archive of previous backups that we don't really need online - particularly now that I have fiber internet - and with the crappy COVID-affected exchange rates is costing me more than I really want to spend for something that is a minor convenience.
I also have backups on the new server and on a VPS here in Sydney, and will be moving to encrypted backups on Backblaze B2 now that they are S3 compatible, so that dedicated backup server is really becoming superfluous.
- I've been looking at a company called SSDNodes for a while. They offer some very good deals on virtual servers - very good deals - if you are willing to pay for 12 or 36 months in advance.
So what's the catch?
Well, they're not a fly-by-night operation, or if they are they're not very good at it; they've been around since 2011. Multiple reviews say they they oversell CPU (and possibly memory, but definitely CPU) - that is, when you pay for a four core VPS you don't get four cores all to yourself.
Now, that is normal for cheap virtual servers. If you get an entry-level Lightsail node from Amazon you get one core (because you can't have less) but if you run a sustained load it drops to an effective speed of around 100MHz.
Amazon don't oversell, as such. Instead they have a system that monitors every single virtual server second-by-second and deliberately chokes the ones using too much CPU time or disk I/O. But smaller operations like SSDNodes don't have that level of control.
SSDNodes themselves say that their servers aren't overloaded - across all their datacenters they run at an average of 40% load and they balance new VPS creation to prevent hotspots.
Is someone lying here?
No. In fact, this is exactly what I'd expect.
SSDNodes run on Intel E5 and Gold series Xeon servers, very common and perfectly normal. Those processors have hyperthreading to improve multi-threaded throughput and turbo boost to improve single-threaded latency.
But the combination of those features completely screws up Linux CPU load figures.
Linux counts each hardware thread as a standalone core for load percentages, but it's smart enough to allocate to individual cores before using the secondary threads. Since hyperthreading gives - if you're lucky - a 20% performance boost, that means that when you have all cores active - and are seeing a 50% CPU load in your monitoring tool - you are really running at 80% or more of your CPU capacity.
As an added bonus, turbo boost will help increase clock speeds on a lightly loaded CPU, but will spin down as more cores become active. That means that a server at 40% load is likely really at 80% of capacity already. And if your true average across all servers is 80%, you will definitely be experiencing hotspots.
So, in short: They are technically not overloaded, but they are pushing the boundary of overselling in a (successful) effort to keep prices down. You will experience cases where your VPS runs slower than you'd like (particularly because self-similarity likely means that your busy times are the same as other users' busy times).
Fine for some use cases, but definitely not for anything real-time like game servers or video streaming. For those you need a dedicated server or a mainstream AWS instance with dedicated cores - either of which will cost quite a bit more.
- Gigabyte's B550 Aorus Master has three PCIe 4.0 x 4 M.2 slots. (Tom's Hardware)
The B550 is an updated X470; it doesn't do PCIe 4.0. So all three of those must come from the CPU, which means that the main PCIe slot is only 4.0 x8 - at least, if you have more than one M.2 slot active.
Which is still enough to run pretty much any graphics card, so it seems like a reasonable tradeoff for a mid-range motherboard.
Or, alternately, the leak could be wrong and the second and third M.2 slots are PCIe 3.0.
- How well does MongoDB handle distributed transactions? (Jepsen)
By default, not well. And if you follow the recommendations rather than the defaults, slowly.
Which is the standard answer for "How well does X handle distributed transactions?" for any X supporting distributed transactions at all. Unless you have a Tandem Nonstop. Which I'm guessing you don't.
- The Mac has now been x86-based longer than either PowerPC or 68000. (Six Colors)
There wasn't much overlap in the PowerPC / x86 switch either.
Posted by: Pixy Misa at
10:19 PM
| Comments (5)
| Add Comment
| Trackbacks (Suck)
Post contains 800 words, total size 6 kb.
Bee And Cabbage Pie Edition
Tech News
- So now I have routing working properly between nested containers on my private cloud. Well, I had it working already, but I had each IP listed as its own route. Fortunately I didn't need a lot of routes so far, but that was going to get out of hand soon.
It makes sense that you have to be very precise with the route command. It doesn't make so much sense that they don't bother to document it.
- TechDirt is drunk again.
- TSMC is building a $12 billion 5nm fab in Arizona. (AnandTech)
Planned to start construction next year and come online in 2024. By then 3nm will be in full production, but it's likely this factory can upgrade to 3nm (and 2nm) as needed.
- ASRock has announced a server motherboard for the Xeon W-1200. (AnandTech)
Not sure how popular it will be with its crazy power draw when ASRock is already shipping Ryzen server boards, but I'm sure someone will buy it.
- Facebook is buying Giphy for $400 million. (Axios)
It's a server with a bunch of images and a copy of Elasticsearch.
- In a surprise move, all social networks have withdrawn business operations from France. (Tech Crunch)
Of course the content is still available and now outside French control, so all France has done is lost is jobs and relevance.
- This code does not do what you think. (Reddit)
If you changetwo
to read as follows:def three():
numbers = range(6)for p in (2, 3):numbers = [ i for i in numbers if i % p ]return numbersThen it does precisely what you would expect. And that's the formulation I always use, which is why I've never tripped over this one.
Explanation here.
Disclaimer: Not that they had a whole lot of either to begin with.
Posted by: Pixy Misa at
12:57 AM
| Comments (8)
| Add Comment
| Trackbacks (Suck)
Post contains 303 words, total size 3 kb.
Friday, May 15
route add -net 10.1.21.0/24 gw 10.1.1.21
You have to specify -net, even though it's obvious from context. You have to specify .0. even though given the context it is clearly irrelevant. And no, they don't document that anywhere.
Posted by: Pixy Misa at
03:51 PM
| No Comments
| Add Comment
| Trackbacks (Suck)
Post contains 87 words, total size 1 kb.
So We Put A Cloud In Your Cloud Edition
Tech News
- Nvidia has announced their Ampere A100 accelerator card thingy. (AnandTech)
This is a 54 billion transistor, 400W, 826mm2 chip built on TSMC's 7nm process, with 40GB of HBM2 memory which is odd becuase the photo shows six HBM2 stacks.
For compute, it offers 50% more single and double precision throughput than the Radeon Pro VII. It's a custom NVLink module, though, not a PCIe card.
- So if you want something you can buy and put straight into your workstation or server, AMD's new Radeon Pro VII that I just mentioned is probably a better bet. (AnandTech)
This is a dual-precision-unlocked Radeon VII, offering nearly twice the DP compute performance at 20% less power.... For 150% more money than the consumer card.
It does have six mini DisplayPort outputs so if you need to run a lot of monitors and a lot of compute and have access to the corporate credit card, maybe not a bad choice.
- Intel has launched their Xeon W-1200 series. (AnandTech)
These are the exact same Comet Lake CPUs with the exact same chipset only the names have been changed and they're not cross-compatible.
I think I said before that these were server processors, but they're actually workstation processors, though some dedicated hosting companies do put them in servers. That means the same 235W power consumption on ten cores. The server models usually try to tame that a little.
- Nvidia also announced their new DGX A100 supercomputer module. (WCCFTech)
This incorporates eight of the A100 compute modules, 320GB total GPU memory, 1TB system memory, eight 200Gb/s InfiniBand/Ethernet ports, six NVMe drives, and dual 64 core AMD Epyc CPUs into a 6U 123kg rack-mount unit.
I'm guessing they needed PCIe 4.0 bandwidth and Intel was late to the party.
Oh, and it uses 6.5kW of power at full load, though I'm not sure how since the GPUs only account for half of that.
- Your laptop may be faster than your server.
If you're deploying to the cloud - whether that's Google, Amazon, Microsoft, or IBM - you're probably getting power-optimised CPUs running at around 2GHz and with very limited turbo boost. Disks are network-attached and bandwidth and IOPS limited.
Your laptop - if it's a recent model - can probably boost to 4GHz or faster and has an NVMe drive capable of a couple of gigabytes per second and hundreds of thousands of IOPS.
So performance on test can be significantly better than in production.
We solved this at my day job by building our own cloud on Threadripper-based servers. Zoom zoom. Three times faster than our old cloud servers at one fifth the cost.
A whole lot of work though.
- Proxmox VE 6.2 is out. (Serve the Home)
I tried out Proxmox VE back when I was trying to deploy Mari. It worked, but didn't solve my networking problems with LXD (which I have now figured out).
I'm minded to create a simple, open-source, LXD management panel that applies everything I've learned.
- Every Wordpress plugin contains critical security flaws. (Bleeping Computer)
If Google can't get it right... Wait, Google is run by idiots these days. I guess this doesn't really prove anything.
- Amazon's new Fire HD 8 is the iPad for the under-$100 crowd. (ZDNet)
It bumps memory from 1.5GB to 2GB - or 3GB on the HD 8 Plus. But it keeps the same old 1280x800 display, which simply isn't enough for reading books.
Also, the Kindle Fire is not only not available from Amazon Australia, Amazon US refuses to ship it to an Australian address.
Why they persist with this fuckery I have no idea. Regular Kindles, Fire TV, Echo, and Ring devices - all the other Amazon electronic devices - are available here. Just not the one worth buying.
- The TCL 10 Pro costs $450. (ZDNet)
6.47" 2340x1080 AMOLED display, 6GB RAM, 128GB storage, 64MP main camera plus untra-wide, macro, and low-light cameras, plus a 24MP selfie camera in a teardrop notch. Headphone jack and microSD slot, and an in-display fingerprint scanner. Oh, and NFC, an FM radio, and an IR port.
All powered by a Snapdragon 675, which is not a high-end part, but is better than Qualcomm's specifications make it sound. They list it as an eight-core Kryo 460, which sounds like one of those cheap eight-core A53 Mediatek parts.
But in fact two of cores are A76 and will deliver fairly solid performance, though still a good 30% slower than a current A77. (And worse on multi-threaded loads, since it only has two fast cores, not four.)
Still, it ticks all the feature boxes and the price isn't bad.
- Deno is node only TypeScript. (ZDNet)
Built in Rust.
Which makes my head hurt.
It does sound better engineered than Node, but since Node is a train wreck where the train is made of burning garbage, that is not hard.
- Google's Pixel 4 is not selling. (Thurrott.com)
This is because the Pixel 4 is not a good device, which is because Google is run by idiots.
- PrintDemon considered harmful. (ZDNet)
It contains a local privilege escalation flaw - in every version of Windows dating back to NT 4.0.
It can't be exploited remotely, so it's not a disaster. If you download and run something nasty on your Windows computer you're in deep trouble even without it stealing admin access.
A patch is available now.
Posted by: Pixy Misa at
01:14 AM
| Comments (5)
| Add Comment
| Trackbacks (Suck)
Post contains 930 words, total size 8 kb.
57 queries taking 0.2142 seconds, 398 records returned.
Powered by Minx 1.1.6c-pink.