Slashdot Mirror
Linux 2.6.20-rc6 Kernel Performance
Michael writes "The Linux 2.6.20 kernel will feature KVM support, Playstation 3 support, and a variety of other improvements. With the Linux 2.6.20-rc6 kernel out the door, Phoronix has written a performance comparison of the Linux 2.6.20-rc6 kernel against the 2.6.19 and 2.6.19.2 kernels in a variety of benchmarks."
Why is playstation support being included in the kernel? Is that really necessary?
Furthermore, the article didn't exactly make it clear what the support is. Can anyone clarify?
Not the most scientific benchmarks in the world
All it takes to get a /. front page link to my ad-word laden website is to create a few bar graphs showing that nothing has changed in the last few kernel revisions and add 2 paragraphs of filler text?
So the bottom line here is that they're almost exactly the same?
In the land of the blind, the one-eyed man is kinky.
Is this a joke?
He's running kernel benchmarks on a laptop?
Looks like half the things he was measuring were I/O bound? On a laptop?
Phoronix.... more like Moronix.
I tried to read the article, but someone has vandalized it with double underlined words all over the place and annoying popups when your mouse slides over them. I closed the window.
Since the article doesn't have any content, I assume this was a badly disguised slashvertisement? None of those are even kernel benchmarks.
/. down already. You're not even trying to compete with sites like Digg are you.
Stop the bullshit ads or just shut
- Adam L. Beberg - The Cosm Project - http://www.mithral.com/
Seems pretty pointless. It'd be more interesting if they threw a few NT kernels in there though...
Yeah, that was a totally worthwhile read, no?
Let me give everyone else the bottom line, and save you two or three minutes of your life, that you'll otherwise never get back: Now, back to our regularly scheduled Slashvertising....
"Ladies and gentlemen, my killbot features Lotus Notes and a machine gun. It is the finest available."
The values are so close, they should have run it many times to at least get a variance and know if those minor differences are "consistent".
Not the most scientific. At least give us a standard deviation or even a range of what the results are. There's a slight difference, but are they statistically significant? Can't tell with the data provided. And given how close the results are, I don't think it matters.
With the Linux 2.6.19 kernel coming out last November and only two additional releases in the 2.6.19 branch, the Linux 2.6.20 kernel is certainly coming quickly. Linus Torvalds had mentioned in the 2.6.20-rc6 release announcement that this is likely the last release candidate. However, even with this quick kernel release coming the features are definitively impressive. Sony Playstation 3 support and Kernel-based Virtual Machine support are among the exciting features in this release. From today's testing in our environment used and set of benchmarks, there were no definitive performance gains or losses seen throughout the set of tests.
It's nice to get features without sacrificing performance. The added PS3 support would nab those ubuntu people to put it on PS3. Not only that, but yellow dog might get some competition if some peoepl decide to make their own PS3-based distro with all kinds of extras.
In Soviet Russia, dots slash you!
It's support for the cell, and some PS3 hardware.
Why would it not be included? If the kernel is still shipping support for 15 year old legacy ISA hardware (yes it is) and Cyrix X86 optimization s(yes it is), it can include support for the PS3, which is likely more in use than either of the above.
There is support for hardware in the kernel that is so obscure that there are probably less than 100 people in the world still using it. There's nothing wrong with this - this is why Open Source beats closed source for overall hardware support - as long as someone is around using it, and someone else maintaining it, there is no reason to remove support for it.
Well it's a release *candidate* so you certainly shouldn't install it on mission critical systems no. The point of an RC is that it's believed to be 'complete' but should still be tested before use in production.
Personally I do not own any gaming consoles, but I do love to play games on my linux machine. If we can make a large gaming society on linux, we may start to get more PC vendors interested in linux, and supporting linux. The more I think about it, the more I think this is a great idea.
Support the source, Open Source! An entire site developed with OSS
Does anyone know how much of the PS3's hardware is actually supported? When you run Linux on Cell, is it actually using all of the Cell cores, or is it just using the main (PPC-like) one?
It seems like Cell is probably going to be a lackluster performer, if only the single main processor is used; at that point it's just like using a 3 or 4 year old PowerPC system. But if Linux can support its additional hardware and coprocessors, it seems like you could do some neat stuff with it; I'd think that you could make a nice media-PC frontend on it, for pushing HD video around.
Seems like getting software to take advantage of it, would require changes both to the kernel, and also to GCC, in order to produce optimized binaries for it, not to mention various pieces of software themselves (rewriting for greater parallelizability).
Still, it's a neat hardware platform (that's about all I have nice to say about it, actually), and it's a good bet that at some time in the future, they'll be available inexpensively on the used market. Anything that starts the process of getting better support now, seems like a good thing to me.
"Ladies and gentlemen, my killbot features Lotus Notes and a machine gun. It is the finest available."
With a great article title like "Linux 2.6.20-rc6 Kernel Performance", I figured there must be something really notable about this release - and yet, there isn't. Summary: some things a neglible amount faster, some a negligible amount slower. If the "firehose" feature I've seen is anything like a preview of how submitted articles are reviewed for publication, I'd say an important part of this process would be at least a review of the linked content to determine whether or not it satisfies any reasonable criteria of newsworthiness.
I'm not so much worried about the "science" behind the benchmarks. I'm still wondering about the point of the benchmarks. I can't think of a more boring set of benchmarks than comparing minor revisions of the same kernel on the same hardware. That is, unless there was some reason to believe that one would be faster. Somehow their "LAME" comparison seems appropriate. "In this set of benchmarks, what we are comparing is LAME."
-matthew
"THERE IS NO JUSTICE, THERE IS ONLY ME." -Death
I think supporting multiple hardware platforms is a good thing, and does not detract from an OS. Especially if the hardware in question has awesome floating point performance. A beefy dual-core processor plus 6 RISC stream processors thrown in? Sign me up. Since the PS3 isn't panning out as a game machine, perhaps Sony should refocus its marketing on the scientific community :) Or port RenderMan to it.
You misunderstand the Linux kernel ethos. The idea is to include *everything* "in the kernel", but you only have to compile the parts that you want. That way there is a central place to track all changes and maintain compatibility and consistency between all parts of the kernel, without having to set an internal interfaces in stone.
It's not "bloat" if it's only in the source. Simply put, you don't have to include PS3 support in your binary version. In other words, the only way it affects you is a few extra bytes to download when you want to compile it.
Well, that's three presentations/impressions per person for each ad. Just because he is a greedy slashvertising asshole doesn't meant that he is a total moron. He is smart enough to know impressions mean money, at least for him if not his advertisers.
Have you ever rolled your own kernel?
:)
You might want to try it sometime as its an educational experiance.
For almost every hardware option you can include whether to build it in, build the module and have the module hooks for it built in or not to include it.
So if you know exactly what hardware you are going to be using you can build a very very small kernel (the theory being it should still be possible to boot a full OS from a floppy, though I haven't tried to do that myself for a while)
If you just want compatability you can build modules for everything, the kernel itself remains fairly small though you do end up with a lot of modules you will likely never use.
Most people go for something inbetween (as do most distros)
About the only way you would get bloat int he kernel is to build everythign built in - and I can't think of a reason anyone would want to do that - ok make that a sensible reason - with geeks the "because I can" argument will always come up
$_="Slashdotter";$syn="OTT";s;..;;;sub _{print shift||$_};s!ash!Perl !;s=$syn=ack=i;tr+LLEd+BLAH+;_"Just Another ";_
No. as kernel developers are concerned, the PS3 is just another computer. A heck of a powerful one, BTW, even if it's a wild beast to tame, but I digress. This has nothing to do with MS bloat, since MS only support one major platform in Windows, while Linux is known to run in a plethora of different archs, and the support must be in the kernel for it to work. I understand PS3 support might not have been in the kernel main tree, and possibly that's what you talking about, but there is no good reason not to; this way basically any distro can provide a PS3 version with less effort, and the support becomes standart instead of consisting of different patches. As I said, to the kernel devs the PS3 is just another arch; this has nothing to do with game support either. The PS3 is capable of a plethora of things, and having Linux available allows you to use a PS3 as a "general purpose computer". If whether this leads public perception to think of Linux as a "gamer OS"... that is another story, but I highly doubt it. Anyway, can't be much worse than being a "hacker OS" right?
Where is that guy who'd die defending what I had to say when I need him?
Given the comments recently about multicore processors being the norm, and the greater push for specialized silicon, optimizing Linux for Cell processors could be a positive step if we see these innovations move into more mainstream use.
Having said that, has anyone benchmarked the performance of the kernel on the Playstation? Is it only at a "it runs" point, or is there actually decent tailoring to the Cell processor, so that it takes advantage of the architecture?
This doesn't have a thing to do with games. It's simply another hardware platform for Linux to run on.
i thought quite the opposite. though, honestly, after seeing a couple of the graphs, i decided not to read the content. comparing application performance on different versions of the kernel seems rather stupid.. like seeing if your car goes faster when you give it nicer seats. consumer application performance is largely dictated by the application code itself and the hardware it runs on. maybe it can handle multi-user I/O and multi-process scheduling better, but I didn't see much of that here.
It's a bit misleading to say that it gives the kernel PS3 support. What it does give is Cell support and other PS3 hardware support. The Cell is made for a gaming machine but the design itself is for a more general use. They even have blade servers with them. This is serious Internet Business we're talking about with support from IBM.
Right now you only get frame buffer support so even playing serious Internet games is not really enjoyable. No GPU support yet.
What sane person is going to think that Linux is a gaming platform off the bat? And what is even wrong with that? Is someone going to take IBM less seriously now that they are supporting CPUs that are specialized for crunching numbers and give you stunning real time visuals?
bah. start over
Kernel developers regularly hunt elusive speed boosts which can only be detected by specialized benchmark. 2% on something as generic as kernel compilation is fantastic.
Of course the tests probably weren't conducted in a sufficiently scientific way, so the measurement error probably swamps the 2% improvement. If it can be independently repeated, congratulations are definitely in order!
Finally! A year of moderation! Ready for 2019?
It's not Microsoft-style bloat, it's Linux-style bloat. It doesn't have any effect on the kernel binary unless you enable it. And you can't enable it in an x86 or x86_64 kernel, because it's a PPC platform configuration. Linux-style bloat is that the kernel source tree has everything possible in it, but it only gets built if you actually manage to ask for it, and then it's generally a module and only gets loaded if you actually need it. But it's in the source tree, because developers want to know about as much as possible of the code that might need to get updated if things change or might have bugs or incorrect assumptions.
This binary is HUGE, the idiot programmer must have put comments ALL THROUGHOUT his code!!
Having all these things "supported" is nice, but I think the idea of constantly piling more and more things into the mainline kernel is A Bad Thing. With the way things are going, it's probably more efficient to download just the files required for a configure script to run, then have /that/ retrieve the things you actually care about, rather than downloading all 9 gigs* of source (*projected for inflation so this post is relevent next week) at once, only to discard most of it.
Actually, that sounds like a good idea. Does anything like that already exist? (I havent bothered to build the kernel in a year or so, but KVM may have me jumping on it.. and I dont know if debian supports my game pad..)
-- 'The' Lord and Master Bitman On High, Master Of All
Tom's Hardware started using them, I was disappointed, as the site is riddled with ads.
However, turning off JavaScript is helpful. Get Web Developer's toolbar for Firefox which allows you to turn such functions on/off on the fly.
Windows supports lots and lots of hardware (different usage of the word "supports lots of hardware" than Linux uses, to be sure). New manufacturers come about and write drivers for their own stuff, and distribute those drivers themselves.
/weird/ to retrieve a driver from a third party. You need to compile something yourself (eek) or hope that it gets included in the mainline.
The Linux model is suprisingly backward for something "Open Source". In a world where customizing things is considered the norm, with drivers it's just plain
But then, why should I download code for Steve's Magic USB Vibrator if I'm not even going to use it?
I dont understand the idea of "Well, if you dont want it, don't build it!"
If I don't want to build it, why should I even get it in the first place?
This is a problem which can easily be fixed, may already have been fixed. But for now (or then), it'[wa]s annoying.
-- 'The' Lord and Master Bitman On High, Master Of All
I can run it on a PS3, but I still can't run it on an Asus A8V motherboard, because there are no working drivers for the onboard SATA controller. Boo.
A slashdotter who didn't build his own computer is like a Jedi who didn't build his own lightsaber.
Can the creators of this technology choose a better word for it? KVM is already widely known as a KVM switch, Keyboard, Video, Mouse. It lets you connect multiple computers to the same monitor,keyboard and mouse and switch between them.
Choosing the same acronym for this new technology is only bound to cause confusion.
AccountKiller
With that kind of performance increase, my Gentoo laptop is going to be screaming along after the release of 2.6.20.
Ok...was that sarcastic enough? With this crowd, one can never quite know.
Yeah. Gamers have Windows. Let them play. They may come when they grow up.
As for the PS3, it's an incredibly fast computer for its price. Of course, it's incredibly hard to program, but we don't do such things because they are easy, right?
http://www.dieblinkenlights.com
It's about time this Linux thing added support for these input/output devices. Even M$DOS has supported these for quite a while. I was getting tired of controlling my lappy by sending brain signals alone and attempting to understand the reply just from the sound of the clicking hard disk.
So, does Linux 2.6.20-rc6 really mean it's a stable release candidate that I can copy onto all of my servers which control the nuclear missles in the continental United States, or is this just a beta and I should wait a week?
This is an RC by the core kernel team . Once the release is final, you should still interpret it as being "released to the distros" for further testing and stabilization. All critical servers should stick to the kernels provided by their distros, unless they can afford extensive testing on their own.
"When I first heard Daydream Nation it quite frankly scared the living shit out of me." -- Matthew Stearns
Gotta ask: will PS3 support in 2.6.20 allow me to turn a PS3 into a MythTV frontend?
"this..."
I've been compiling my own kernels for years. He speaks the truth.
tasks(723) drafts(105) languages(484) examples(29106)
I know you were joking, but actually removing inline functions can make the kernel faster in some cases, and the effect on kernel size is dramatic:
:)
http://lwn.net/Articles/82495/
http://lwn.net/Articles/166172/
And let me say, LWN kicks ass.
People suggested it multiple times already, but it never gets done, probably because it's complicated.
For example, if you actually compile from source, this is undesirable. If you're compiling from source instead of using your distribution's kernel you probably have some specific reason for it, like wanting to apply a patch. Now what if this patch touches some kernel internals (say, the USB subsystem), and updates multiple drivers? Then applying the patch would fail.
Also, having the full source is a VERY good thing if you just replaced your network card, have no net access, and need to build a new kernel to get it. Then what if you didn't download any other network drivers? Bootstrapping problem there too: if you download drivers, how do you download the modem/network card driver? If you ship those, why not others since they're all in the kernel to start with? Making arbitrary distinctions is strange.
If you use binaries, you again have the same problems. If you have all the drivers installed that takes what, 20MB disk space? And in exchange if you swap the network card, a recent distribution will detect the new one and automatically load it, instead of making you mess with burning a driver CD on some other computer.
Then, Windows has exactly the same behavior! When you connect say, USB speakers (those with an integrated soundcard) they just work, because the driver for them was installed with the OS, even though it was unused at the time. It doesn't ask for the CD, or download anything.
Basically, making it download drivers is just more complication, extra things that can go wrong and less user friendliness in exchange for just a few MB of disk space, which doesn't amount to anything noticeable on a disk bought in the last 5 years anyway. Now if you're for some bizarre reason still booting Linux from a floppy, or work on something embedded that runs on 8MB of flash, then this makes sense, but that's a very uncommon situation.
IBM is starting to push out Cell based hardware. That means heavy iron with multiple processors each with multiple types of cores. You know what people are using to learn how to program that? PS3s. You get all the techniques down on a sub $1K chunk of plastic then you migrate those techniques up to $1M+ big iron.
I think if you look at the people contributing to the Cell frameworks & libraries, you're going to find IBM investing heavily in time & lines.
It's slightly faster.... TO THE MAX!
If I don't want to build it, why should I even get it in the first place?
That's an interesting question, that I believe will come up more and more in the future as the amount of hardware and drivers increases. For now, I think the main reason for the current model is that it makes things simple for the kernel developers, with a minimal impact on the users-- a few extra megabytes of source code. (Users here being the people who compile their own kernels.)
I personally like this distribution model because it is very enriching. I can't count the number of times I went browsing kernel source looking for code examples or designs, and found something interesting in some obscure driver I would never actively go get otherwise. This is very similar to how many Linux distros come with an enormous mountain of applications that cover most things one would ever want to do with a computer, and it gives me warm fuzzies.
It is a very different experience from how commercial OS works. Using any new Windows installation for me always came with an uncomfortable nagging fear of the "feature wall;" there was inevitably something interesting I would want to do with the system that doesn't come with the OS, and costs money. Want to actually modify this Word document? Buy Office. Want to *gasp* compile your own programs? $$$ Visual Basic, Studio, etc. Want to edit photos? Buy photoshop. Oh, you wanted to make drop shadows? You'll also need this random Plugin that costs extra. It always felt like a fundraising dinner or movie theater with their expensive popcorn. I still have bad memories of college where we always fought for a seat at the non-lame Sun workstation that came with cc.
Tsunami -- You can't bring a good wave down!
On 2.6.19-ck1, I ran into a number of show-stopping problems with I/O resulting in several severe hangs (which I never bothered to debug) which I have linked to I/O. I've been hearing reports from other users of the -ck patchset that they are experiencing similar problems with 2.6.20-rc5/6.
I think I'll wait a bit and stick with 2.6.18 for now.
grey wolf
LET FORTRAN DIE!
I've played with PS3 linux. I can tell you, the hypervisor is just that. It virtualizes the PS3 hardware. ...
I don't know about anybody else, but I find this just conceptually fascinating. Where does the hypervisor run from, anyway? Is it in the machine's ROM, so that there's no way to prevent it from booting? (Without irreversible hardware modifications.)
I was just wondering whether it's possible to get rid of it, and boot Linux on the bare metal, or whether the hypervisor is tied into the hardware so tightly, it's impossible to remove and install a new Domain 0 operating system.
Getting Linux to run on the bare metal, 'below' the hypervisor, will be an interesting exercise in what I suspect may be a large part of the future of "unauthorized" computing. I don't think it'll be long before most consumer systems have something like that in place, so it'll be a good intellectual challenge, if nothing else, to see if it can be gotten around.
"Ladies and gentlemen, my killbot features Lotus Notes and a machine gun. It is the finest available."
If your gamepad uses USB and is HID spec compliant then Linux should have no problem. It was literally silent plug and play with no configuration necessary in both DOSBOX and ZSNES for my Logitech gamepad and joystick, all I had to do was compile in the right modules.
I don't think kernel download size will outpace technology (or even the sense of efficiency). In 1993 or even 1998 downloading 40.8MB of source might have been a bit excessive.
Still I find what you are proposing to be intriguing and I wish to subscribe to your newsletter. Basically you mean something like an apt-kernel-config tool?
I may make you feel, but I can't make you think.
Hardware support belongs in the kernel. We're talking about another processor architecture which has historically always been shoved in. Would someone ask why MIPS support was added or why any number of other architectures was added that are rarely used (compared to x86)?
/proc. That is something I'd be more interested in. While the cell proc has some great features, I don't see a platform that I like yet. (Sorry a game box isn't a solid platform for the computing I want to do.)
Now the better question is if someone can/will make a standardized PC with the Cell
I also bought a PS3 two weeks ago primarily to develop on the Cell processor. So far, I have written a prototype application that runs 5x faster on the PS3's Cell than on the highest end Woodcrest Xeon at 3.0 GHz. When I am not coding on it, I am also using it as a Blu-ray player. Given all this, plus the fact that the PS3 is a next-gen console (though I don't plan to buy any game), I am probably one of the few to recognize that $500 is dirty cheap for such a polyvant device !
Given that the article, obstensibly about kernel performance, makes no effort to actually cover kernel performance...
I was suckered in; it looked like a worthy topic. Now, I ask myself: Why did I bother?
Just another "Cubible(sic) Joe" 2 17 3061
When it booted, two big Tux penguins were displayed indicating dual CPUs. Then after a bit of booting six more smaller Tux penguins appeared beneath the first two. So I suspect all eight cores are in use.
Well, hell -- eight penguins? I know that's all the benchmarks I need. Now, where's my credit card?
"Ladies and gentlemen, my killbot features Lotus Notes and a machine gun. It is the finest available."
That's not true, the code must have declarations to evaluate if we have XYZ feature loaded wich DOES impact in the performance
Those code declarations are called #defines, and the C preproscessor automatically removes them, the C compiler itself never even sees them
YHBT, I think.
(I'm hoping the GP was a troll...I've always assumed that people like that were too stupid to use anything other than phpBB-based boards.)
"Ladies and gentlemen, my killbot features Lotus Notes and a machine gun. It is the finest available."
Some of the figures went up by ~1%! Did they move something out of the kernel?
He's running kernel benchmarks on a laptop? Looks like half the things he was measuring were I/O bound? On a laptop?
Point well taken, but I have no problem with this.
At least the notebook computer was built by one manufacturer, who (presumably) ensured that all the components played well with each other. Other than that, I don't care what machine he ran it on - an AST Pentium I would have been fine with me.
The notebook may be slower than a comparable prefab desktop or server, but we still had A/B comparisons taken on the same hardware.
Despite the slower IO, it also more accurately simulates bottlenecks seen in older hardware, or hardware where the BIOS has been "tweaked" by a "tuner" who knows nothing about what he's doing (we've all seen hardware like this, my favorite was back in the days of the 486, someone who wasn't smart enough to know how to buy jeans which fit properly had disabled the cache because it "slowed him down").
Remember, we have an A/B comparison. Performance differences between kernel versions should scale by the same linear coefficient as the comparison between this notebook and whatever else you plan on running on.
And furthermore, notebooks are a very common platform today, and their usage does reflect a real world condition under which any new kernel will have to operate.
Fire and Meat. Yummy.
Basically you mean something like an apt-kernel-config tool?
Why qualify it? I mean I want to type$ apt-kernel config
It's not actually at the point where it matters for the most part yet, or at least it wasn't last time I built a kernel (a year or so ago), but it's the principle of the thing more than anything else.
Of course, some of the most numerous drivers are network drivers... not the kind of thing you want to be picking and choosing piecemeal while reliant on a network source to get at them.
-- 'The' Lord and Master Bitman On High, Master Of All
Although I agree with you in principle, KVM switches are obsolete anyway.
You want a KVAM (pronounced to rhyme with "shazam!" for marketing purposes) switch.
Keyboard, Video, Audio, Mouse.
I put a standard Fedora Core install on a machine with one of those motherboards. I didn't notice anything unusual at all during the install and the user has had no problems with it in the months it has been in use. I can't remember what it had to start with (it was the kernel on the install CDs) but it now has kernel 2.6.17 - which matches the new Fedora Core6. I know old knoppix disks with old kernels don't have the nforce driver but linux distributions less than a year or two old should work with your SATA controller.
Good point - from your comments above you fill that role yourself perfectly.
To be serious - the idea is you don't have to hunt for drivers from weird spyware ridden corners of the net. The people who put together your distro will have most likely compiled the kernel modules for it (all those bits you don't want but might someday) unless it is new, in which case you need to download an updated binary kernel and binary modules or compile it yourself and only get the bits you do want.
I wouldn't worry too much about it. KVM is not going to be championed by the likes of Redhat, SuSE or any other mainstream distro, so you won't really see or hear much about it in the real world. Their VM of choice is Xen and with good reason. Xen Paravirtualization narrows the gap between native and virtualized performance, and Xen's killer feature, Live Migration, turns the world of HA clustering on its head.
You get:
1 disabled SPE
7 working SPEs (each has 128 128-bit registers)
2 PowerPC 64-bit w/ vector units (one PPE is two-way SMT Power 970)
That is 9 CPUs total, 7 weird and 2 normal. Linux mainly runs on the 64-bit PPE cores, but there are ways to run code on the SPE cores.
Decent SPE documentation is nowhere to be found AFAIK. Perhaps it can be derived from a gcc port.
it's good to include an error calculation. The first page of benchmarks could really be sampled from the same population (effectively), but its hard to say. I think for many of the tests shown, there would be variation.
If this were done, it would make the interpretation more straight-forward.
The Processor is about equivalent to a mid-to-late Pentium III.
No, it's actually about the same as a Powermac G5 at the moment. Read more here.
"Results
Overall Score
PlayStation 3
105.2
Power Mac G5
106.9"
I gotta compile this for my 970MP with 8GB of RAM. I just wish Nvidia would get thier head ouf thier ass and open source the driver since I need to compile it for my powermac G5. other than that Linux is much faster than os/x. And with the new
X server enhancements, damn I won't miss OS/X much. gotta love the new 3D desktop. It's awesome on my AMD Linux box.
I gotta get it on my G5, then I can sell my AMD box and put more memory in my G5. Or save up for a quad AMD box.