Like Rabbits

XenServer performance tweaks!

2009-09-11T22:30:00.005-05:00

Just a few notes I'm making here about performance tuning on XenServer 5.5.0 - this is pretty much unsupported by Citrix, but since it's "just" Linux and if you have enough nous, it does make a difference.

To get your NICs bonded using an industry standard link aggregation protocol (like LACP), drill down to /etc/modprobe.d/bonding and modify as follows:

options bonding mode=802.3ad miimon=100 downdelay=200 updelay=200

(I wish XenServer used a more modern kernel then I could specify ad_select=bandwidth xmit_hash_policy=layer2+3 too)

This pretty much makes your NIC bond work the same way your switch does, instead of using Linux's own bonding system. For our Cisco switch this works better. You can also go in and edit interface-reconfigure and supply your own bonding defaults (perhaps a little nicer, you can also fix your MTU here without unmanaging your NICs), or try and set these as "other-config" options but that's just a pain in the backside to do for every VM.

Want better performance out of NFS SR's, have an NFS4-capable SAN? Drill down to /opt/xensource/sm/nfs.py and modify the line 64:

options = "soft,timeo=%d,retrains=%d,proto=%s,noac" % (SOFTMOUNT_TIMEOUT,

(the difference being the protocol is prefixed with proto=) then line 69:

util.pread(["mount.nfs4", "%s:%s"

(the difference simply being that the script calls mount.nfs4 instead of mount.nfs)

And that's it. We reduced latency by a couple milliseconds, and got a ton of extra bandwidth out of it (~380MB/s instead of 130MB/s) versus NFSv3. That's before enabling jumbo frames.

Geit PowerDev Meeting Webcam

2009-04-11T10:23:00.002-05:00

Power Developer 2.0.1.1.2.9.99.999.40.19

2008-06-04T05:11:00.002-05:00

Edit 26th November 2008: there is so much to do this is yet again on the back burner, there is a lot of work to do on the power2people site and other work which takes up most of the day. One day there will be a new site (check out the work done by the "new web team" on the Genesi USA main site) but not right now.

A year later and I'm still working at this, although Projects and a lot of the front page have been reworked a lot, I'm working on a real design document to integrate EVERY user suggestion (blog syndication & comments, wiki-style pages - including a forum that supports wiki formatting, the long-awaited RCS integration (I still can't decide if Subversion or Mercurial is best for this) and a bunch of other stuff) and hopefully making things more of a coherent whole and easier to update anything and everything.

Space: 2099?

2008-06-04T04:58:00.003-05:00

Working at home has some advantages, like flexible hours, but the one that's keeping me amused right now is coding with Space: 1999 in the background (gotta love the digital channel revolution in the UK, 30 public TV channels with little on but re-runs of 70's shows :)

I wish they'd remake this. After Battlestar Galactica comes to a close this year, we're going to be watching re-runs of THAT. I know Ronald D. Moore and the Battlestar show-running team are fans of Space: 1999 and I am sure they would jump at the chance; and why not?

It was sci-fi with real intelligence that inspired the story arcs and clever episodes of Deep Space 9, and the new Battlestar (and was depressingly absent from things like Star Trek Voyager, Andromeda, and even Stargate SG-1 etc.) Rethinks of the best episodes with new ideas, plonk on a Bear McCreary reimagining of Barry Gray and Derek Wadsworth's show themes, updated special effects, relevent social and political commentary (Battlestar's handling of terrorism, suicide bombing, prisoner treatment and military justice all around the events in Iraq and Afghanistan, and the obvious election fixing hubub all went to make the show accessible)

I wonder though, since naturalistic science fiction is all the rage these days (Battlestar's wired 1960's-retro-style handset phones and lack of aliens are a another great example), how would you justify sending the moon out of Earth's orbit? Somehow someone has to gel the fantastic of the 60's and 70's with the realistic (and based on real science) of today.

Oh yeah and it has to have Juliet Landau in it. I mean she's the daughter of the two stars of the original, and I bet she could lend a decent hand to it :)

Here's hoping anyway. Next blogs will be about work, I promise :)

Dragging the 20th Century Boy into the 21st..

2007-04-01T22:02:00.001-05:00

Coming to a browser near you;

We've changed the sites we run a lot over the years but each time we have pretty much steered clear of modern browser technology since we have been trying to support operating systems like MorphOS and the lower-end browsers on all platforms (Dillo on Linux is a good example) by not using too much CSS and Javascript.

Since MorphOS has a KHTML/WebCore (that's 'like Safari' to you and me) browser in the late stages of development, there's little point now in keeping a pretty site bogged down in nested tables, hardcoded image backgrounds, invalid HTML and nasty Netscape 3.x tricks.

So, since last week along with the other site updates we've been working on, I've been poking around with CSS, reducing the code footprint of the content engine and adding all kinds of neat stylesheets which improve the look and feel of the website and hopefully make it a more comfortable place to visit. One of these changes will - hopefully - be upgrading the forum to PHPBB 3 for a more pleasant forum experience.

The picture doesn't show much; but it's the changes behind the scenes that make all the difference. The page loads significantly faster - after all, there is less of it. There is less code dynamically generating layout and a lot less (rather pointless theming) imagery to stall your browser. Logic in the code which rudimentarily spaced things out to make them easy to read are now in a stylesheet - which is ostensibly compressed, and cached at your end, so you don't have to wait for that. There are less tables so the page renders faster, too.

I'm pretty excited about doing it, it has been a hard few years being a webdesigner that could not use any heavy CSS. The step after getting Power Developer squared away will be the rest of the Genesi web presence - the store most importantly. We will leave MorphZone and the MorphOS Developer Connection until last just because I want Sputnik to be out of beta for real. That is not to say that MorphOS gets a back seat! We're just taking it one step at a time.. :)

The case of..

2007-01-24T06:05:00.000-06:00

We have our first production case, comes in many colours, but this blue is really quite nice :)

http://www.blastwave.org/dclarke/blog/?q=node/29

Thanks Dennis.

Efika 5200B Galleries

2006-12-23T05:29:00.000-06:00

One gallery for Power Supplies - Sven Luther of Genesi took these pictures to show off his new PicoPSU along with his new Efika.

Another for cases - this is a rather ingenius case made out of the box the Efika came in. If this isn't "green" I don't know what is. It's possible mainly because the entire system even with the graphics card generates very little heat.

You can also make a case out of any 5.25" drive case (for USB, SCSI or so on) as the Efika has been designed specifically to fit in the same space as a standard half-height CDROM. These chassis are usually very cheap if you buy them without the drive, in fact sometimes cheaper even than PC cases.
Why not buy two, and a cheap DVD drive, so your Efika stack can read DVDs via the USB port?

Developers, Projects, Plans..

2006-11-26T07:35:00.000-06:00

The Efika 5200B program is in full swing; boards are going to go out next week to everyone who has an approved project (some are still under review but we will finish this up before the final ones are shipped). As said before we have collated some basic information and all the relevant documentation on the board that will be useful to developers. This support will expand heavily in the next few weeks, based on developer feedback.

There are some great projects in there, and I hope we will see some highly innovative uses for the Efika board and Power Architecture in general.

Once the boards are out there, the roadmap is to get the features everyone needs to fulfil their requirements - grouping developers with similar projects to work together and Collaborate, Subversion and Git repositories for developers who want it, email aliasing, pretty HTML mails for commits and mailing lists for groups of developers (along with the forums). We don't plan to replace such resources as Berlios or Sourceforge (after all they do a great job) but to provide a simple interface which is neither intrusive nor overcomplex. All we ask is the developers update their "Project Blog" with their progress, after all Freescale, IBM and other partners and interested parties check the site all the time and may be more than a little interested in where you guys are going with your ideas.

The next step; provide the dual-core 8641D boards, and do something with the PlayStation 3! I think it is quite exciting what Sony have done with their "open platform", it marks a significant departure from the previous strict, proprietary nature of the traditional games console, but unfortunately a lot of the guys who need to do the work in bringing up distributions are in Europe and they have to wait until March to get their hardware.

I think we at Genesi are well placed to handle an effort centralising and supporting the development of PlayStation Linux and other operating systems. We have already started to set things up to enable these users and developers. A lot more is coming, but at the end of the day it should be up to YOU guys to tell us what YOU want out of it. Tutorials are a good start. We are collating every snippet of information on Cell and PS3 we can and it will go onto the site.

Maybe we could provide easy access to the system (perhaps a few online for 'free' use through the Projects program) and we will definitely be providing more support for the existing Cell and PS3 development community. I would love to know what the community thinks of this, and what they actually require - beyond all the previous discussions we have had on PowerDeveloper and Power.org forums, especially as that was all well before the release of PS3 or any consumer-level Cell hardware.

Have fun, guys, and good luck with those projects!

Efika 5200B rolls out...

2006-11-09T08:00:00.000-06:00

Busy month.

Power Developer is live! We hope this will become the best central (actually it pretty much already is the defacto standard) resource for Power Architecture development, focussing mainly on Genesi platforms. But we like Cell, and the rest of the hardware out there too. If enough people want it, it will go on there.

The Efika 5200B product is on sale and ready for action..

We put up a page hopefully to ease some of the hundreds of emails we get asking for details and documentation on our platforms. For now there is just the Efika version, but we will expand it with the rest (Pegasos) later this month.

The Hard Cell

2006-09-14T09:11:00.000-05:00

IBM, Sony and Toshiba's "Cell Broadband Engine" processor is a really neat thing.

The Cell processor basically consists of a Power Architecture™ processor unit (PPE or Power Processor Element) and 8 SPE units (Synergistic Processing Elements) - these 8 units are essentially their own processor and memory controller, with an insane amount of memory bandwidth (XDR RAM is recommended for it's 8.0 Gigabyte/s sustained transfer rates) and geared up for heavy single-precision mathematics the same way Intel's SSE, Freescale's AltiVec (and IBM's VMX), and any commercially available DSP does. Except this one runs at around 3.2GHz, it has 7 friends, and there is a dual-threaded 3.2GHz Power processor (apparently very similar to a G5) sitting there with dual AltiVec units too (oh my word, and 2 parallel 64-bit FPU units if you believe it..). This chip has a LOT of power.

It's most famous, of course for its use in the Playstation 3 games console (to be released later this year, or early next year, depending on where you are). That is all a lot of people have heard about it. Just recently, though, it is starting to pick up and it is being implemented in many, many other devices.

About this time last year, Mercury Computer Systems announced that they would be using Cell in medical imaging and embedded applications. Their first product was a dual-Cell blade, with a lot of special Cell-optimized imaging software on board, along with a nice big box to test out your new blades in. They didn't disclose a price. But the blades have 2.4GHz Cell processors in them.

They also have a PCI Express card with a Cell processor on it - ostensibly for evaluation of the processor without buying a 470lb system to do it. We do know this board is $7,999 and has a 2.8Ghz Cell processor in it, 1GB of XDR RAM (this is fast stuff!) and 4GB of DDR2 RAM, and takes up two slots of space, and one of your PCI Express x16 slots in a system. You can see it in the picture :)

Today, IBM announced the release of their QS20 blade. This looks identical to the Mercury one, but it has 3.2GHz Cell processors in it. This is effectively the same performance chip that the Playstation 3 will be using. IBM did not release a price, but the rumor is that each blade costs upwards of $18,000. We could assume from this that the original Mercury blade cost about the same.

Then we get back to the Sony Playstation 3, which they say will cost between $500 and $700.

(I am going to mention that IBM also have a free, open-source Cell Simulator which will allow you to play with Cell code in a window on your Windows PC or Linux box. However.. think about how much power (maybe that should be Power..) a Cell chip has, and what a hard time your dinky PC will have trying to mimick it)

The price difference is pretty huge. But the reason is - to cut costs, Sony are using chips which have 7 SPE units available. If one fails in manufacturing, it's still good enough for the Playstation 3. They have also decided that they will only use 256MB of XDR, and 256MB of GDDR3 RAM (cheaper and almost as fast). That's fine. This helps when you need to sell a ~$500 games console and gives more than enough performance for the next 5-10 years of the Playstation 3's lifetime. Nobody would buy a $8000 box to play Metal Gear Solid 4. Plus you do not get high-end medical imaging software with your Playstation 3 - you get Metal Gear Solid 4 and Katamari Insane Asylum or whatever, and they cost $60 each.

However for a software developer, this price difference isn't much use. You get a games console with this great processor in it, but you don't get to use it unless you are authorized to do so by Sony Computer Entertainment in your part of the world anyway. So if you wanted to develop for Cell, you are stuck with the cheapest option: $8000 Mercury Cell card, and you get all this high-end imaging software that you don't want.

I got thinking. That is a pretty high entry point. You aren't going to get many bedroom programmers finding $8000 out of their student loans to get into Cell development. Maybe they will get lucky and the University they are at will have one or two farmed out for student use, though. Maybe not. I think this is going to hurt Cell in the end, and companies who want to use Cell, because its not like they can advertise in a newspaper for "a software engineer familar with Power Architecture, specifically the Cell Broadband Engine".

And what of the actual Mercury card? It is a double-width (takes up two slots), full length, 16x PCI Express device. It's not been in the last 6 months that you have been able to buy (by virtue of nVidia SLI and ATI CrossFire solutions) a PC motherboard with 16x PCI slots so that you could build a machine with a decent graphics card *and* your Cell development card. Buying a brand new PC motherboard, or giving up the graphics board you bought for your PC, is a little far out for me at least.

And what of the use of Cell? I am talking about this processor like it is the best thing since sliced bread (it is :) but you might say that apart from medical imaging it has no real use, so $8000, $18000, bundled Cell APIs for medical imaging and military use is okay. Sure. You can think that. But, we can look at the PC market today and we can see there are lots of really quite strange solutions for really quite strange problems;

Ageia have produced the PhysX PCI card. This board offloads physics calculations from the host processor, ideally for gamers to get that few frames per second more out of their "rigs". It costs about $250, has 128MB of RAM on it, and a 500MHz processing engine.

Bigfoot Networks have just released their KillerNIC network card. This is a $280 Gigabit Ethernet card with a 400MHz processing engine on it. Apparently this will offload your machine enough to reduce your ping times and handle gaming better.

ATI Technologies, Inc. - our great graphics-chip producing friends, have decided that you don't need a Physics processor, because you can reappropriate a standard ATI GPU into doing Physics calculations. Their recommended system for this "Boundless Physics" is an ATI CrossFire setup with an EXTRA card in it - that's THREE of at least a Radeon X1x00 card. They are about $100 a pop so it's anywhere between $100 (to just have the graphics you need plus a physics offload) and $200 (for CrossFire + Physics) to get the ultimate gaming system.

nVidia have also announced support for HAVOK Physics engine in their GPUs in SLI configurations.

And now AISeek have announced their "Initia" AI CPU - another PCI offload card which calculates all those line-of-sight and node mapping formulae that everyone in gaming (and other simulation) so that your main CPU doesn't have to. I have no idea how much this board will cost, but you can bet it will be about $250 again.

How much money and how many slots do they think I have in a PC? Now I need three graphics cards, a physics card, an AI card, a 400MHz network offload card, will I need one to make coffee for me too? It seems I am not alone in being slightly annoyed, perplexed, maybe just plain astounded at the state the industry is getting in here;

http://www.atomicmpc.com.au/forums.asp?s=2&c=6&t=10696

Couldn't I just throw a single Cell processor with a moderate amount of RAM in one slot, and have ONE of the SPE's handle EACH of these esoteric offload functions? I don't see why it couldn't. Each SPE runs at 6 times the clock of these dedicated chips, and it could have up to 8 of them in there.

But back to cost; maybe the cost is too high if you need the 8 SPE one, and a lot of XDR memory. If Sony can get it for less than $500 by removing an SPE, only having half their RAM as XDR, maybe knocking back some more SPEs and using DDR instead can make it even cheaper without too much of a performance loss. After all. If you wanted to offload network processing, AI, Physics and even help the GPU do texture calculations, you would only need 4. And that's before you even use the PPE.

Questions:

How many SPEs does a bedroom/low-need programmer really need for their application for an effective Cell processor solution? 6? 5? As little as 3?

How much cheaper would a 3-SPE Cell be compared to Sony's 7-SPE Cell or IBM/Mercury's super-high-end 8-SPE Cell?

How fast does a Cell need to run to be useful in an application? We can see solutions from 2.4GHz to 3.2GHz on the market. Does this affect how many SPE units you would want to use?

If you could get a $350-$450 Cell card on PCI Express (but not requiring a 16x slot) or PCI, with some ~512MB memory (maybe just a slot, DDR2 PC2-8000 is fairly cheap and has adequate bandwidth at about a tenth of the price of XDR) and a Gigabit Ethernet port (maybe a serial port too) would that be exciting to developers?

We have seen from programs like IBM's Linux On Power competition, Google Summer Of Code, Genesi/Freescale's Pegasos and Efika developer programs, that getting developers to develop for Power Architecture processors is not hard; they are even very eager. Linux support for Cell is supported by IBM themselves and the release fo kernel 2.6.16 (6 months ago) marked the official mainline support of the Cell chip.

If a device like this existed it would be the perfect entry point to both Power and Cell programming for everyone in the world, and as a side note, could completely turn the industry for these "offload accelerators" in gaming on their head.

Final question: Why isn't anyone doing it? Do IBM or Power.org want to reduce the barrier for entry to Power Architecture and specifically the Cell processor in this way?

Full Throttle

2006-06-16T11:28:00.000-05:00

I always used the analogy "using a jet engine to cool it" when talking about Intel servers and the latest big-iron technology. And now..

http://www.technologyreview.com/read_article.aspx?id=16992&ch=infotech

Accelerating Climate Change (Prediction)

2006-02-20T06:08:00.000-06:00

http://www.bbc.co.uk/sn/hottopics/climatechange/

Is it just me, or is using spare CPU cycles - and therefore more electricity, and more heat generated by the systems - not going to do much for the environment and therefore the climate?

Did they just not think quite so clearly?

There are already concerns that computers and TVs are being left on standby and wasting millions of kilowatts per year of electricity, which is contributing to the high cost of power, reduction of fossil fuel reserves, contributing to climate change and so on.

Wrangler, Levi, whatever takes your fancy..

2005-11-30T00:57:00.000-06:00

This post is part 3 of a bunch of posts. Part One was here and Part Two was here

Not in the market for a NAS box? Don't want to store your music on it and play them via the audio out? Well, let's look at a more high-end market. Supercomputing.

"Supercomputing??", I hear you scream in dismay, "but you can't supercompute with a sub-GHz processor with no L2 cache!!!".

IBM begs to differ.

The current world leading supercomputer is BlueGene/L - a monster that IBM have created by linking together 130,000 or so embedded PowerPC processors. Each node has two 700MHz 440-series PowerPC SoC's, and you can fit 64 of these nodes in a "rack". They are so low-power (averaging a watt or two, each node, at full use) that they can be packed in very high density, giving a very very high computing power per square foot of used space, and don't require a personal nuclear power plant to operate. The exact specifications of a BlueGene solution are at the IBM website. It should be noted that the #2 fastest supercomputer in the world is in fact another BlueGene, which simply has less processors in it's configuration. The BlueGene experimental prototype ran 8192 500MHz PowerPC processors and is STILL ranked as #22 (above it are yet 3 other BlueGene solutions).

I am not saying someone will buy 200,000 EFIKA boards just so they can get to #1 in the supercomputing list, although I certainly will not be pulling them away from signing that cheque. But the possibility for a very cheap, very effective cluster solution presents itself. There are lots of clusters in the world at universities and laboratories which have been built by students and scientists alike dumpster-diving for parts. An old Compaq 386 here, an RM 486 there, and the occasional but rare Pentium II perhaps at such heady clock rates as 266MHz.

Why not get out from that dumpster, put your white coats in the laundry and use EFIKA instead?

The MPC5200 is efficient; 400MHz of computing power, at around 1W. No fans. No noise. No ugly beige ATX cases and searching for ancient ZIP RAM modules to put in them. In fact the EFIKA measures about 15cm by 10cm so instead of taking up the side wall of your office building your makeshift cluster, you can simply clear a pile of papers off your desk and use that space.

Black, like my mood

2005-11-30T00:50:00.000-06:00

This post is part 2 of a bunch of posts. Part One is here

Let's look at something technically very similar to the EFIKA which is already on the market - the Revolution KuroBox. This is in fact an embedded PowerPC board using similar Freescale SoC technology at 200MHz. It's primary application is parent company Buffalo's LinkStation NAS product. They ship this box with a 160GB or larger disk inside, and it shares it across the network.

The KuroBox ships diskless as the "tinkerer" version of the LinkStation and allows your average Linux geek to take this board and make whatever he likes out of it. Most applications are NAS (like the LinkStation), web server/ftp, webcam control (using the USB port on the device) and internet firewall/routing.

(Revolution have a second KuroBox "HG" which has a 266MHz processor, 64MB more RAM. The price is $160 for the standard or $240 for the HG model, as I write this)

KuroBox has a few quirks though; flashing the firmware is the primary method of updating the custom Linux kernel, via some apparently quite flakey firmware tools which for a while only worked under Windows 98. You can set it up manually with any OS, though, if you like.

The Linux kernel - as mentioned before - is custom. It isn't a mainline kernel and it isn't even very recent (as of writing, they use 2.4.17). What happens if you build a kernel using the crosscompilers available under Windows (Cygwin, MingW32) or Linux, you flash it to the device, and it happens to be broken? Oh.. well then your KuroBox is broken too and you need to get it repaired and reflashed by someone qualified to do so. If you live in Austin, Texas, you can walk to the Revolution offices and do this. Lucky Austinites! Otherwise, the serial interface - which would be your primary communication with the board - is 6-pin, not officially documented, and apparently NOT directly compatible with RS232 (it requires some extra resistors and so on to make a cable which you can plug into your PC, Mac or PegasosPPC).

So. On your left you have a KuroBox. Nice little machines with lots of articles about booting special versions of Gentoo using it for robotics and x86 to PowerPC migration for embedded work. All the quirks as above, and more that you may find.

I'll stop bashing Buffalo, Revolution and the Linkstation now, and say that on your right you have an EFIKA 5K2 which has more features (audio and a standard serial port), is more expandable (PCI and AGP!), faster (466MHz over 266MHz max), an immensely better development environment and ability to run standard Linux distributions and mainline kernels (2.6.14 worked on release and with GCC installed you can compile this self-hosted and bootable from disk or network) and therefore better value for money if you eschew the pretty black shiny case we don't have right now.

Which would you choose?

Effectiveness

2005-11-29T23:08:00.000-06:00

Advertising using my blog, ohmygosh :)

We at Genesi just released our latest product, something which you may have already seen online before or at the Freescale Technology Forum show in Orlando this year. It is the EFIKA 5K2 (for the insane name-dismantling people out there, efika is Esperanto for "effective", and 5K2 is because it is based on the Freescale MPC5200B processor)

The product page is here: http://www.genesippc.com/efika.php

I guess we hoped we did a good enough job with the EFIKA product page to get across exactly what the board is for - but there seem to be a lot of questions about it, centering around the fact that people don't understand what specific market it is aimed at.

Revelation: once size fits all! Who needs a specific market?

Basically it was designed as a performance evaluation for the MPC5200 processor, with a view to deriving further devices from this simple reference design.

(Parts of the EFIKA are, in turn, derived from the PegasosPPC which is what you might call our first reference design. This design is in fact being opened up and if you are a Power.org member you will soon be able to download it for free)

The EFIKA alone, however, stands out as a very useful little device on it's own. It has a PowerPC SoC (System-on-Chip) with the "e300" core, which is what Freescale are calling the 603e these days. This runs at anywhere from 200MHz to 466MHz depending on what the customer wants.

This chip contains this processor, plus a 100Mbit network interface, 2 USB ports (12Mbit), serial UART (9-pin "d-sub" RS232 on the board), and a DMA-capable IDE interface (44-pin on the board) amongst other things. There are solder pads on the design for 512MB of RAM - we solder 128MB for the current designs.

There is the trusty Sigmatel STAC 9766 codec as used on the PegasosPPC, and stereo audio, microphone and line-in 3.5mm jacks on the back of the board. Additional audio input/output connections on the board (all "Sony CDROM"-style connectors) for TV card, headphone and so on are available.

Last but not least is a full PCI slot. This can take a riser in order to have a PCI slot parallel to the board (rather than at 90 degrees) and this is easily expanded with an AGP riser option - that will mean you can attach any 3.3V AGP card and it will operate as per the AGP standard in "PCI mode". The PegasosPPC AGP port works the same way. Contrary to popular belief this isn't some horrible performance-killing feature. Game frame rates stay much the same. Texture uploads will be slower but most games try their hardest not to do this during gameplay (since even with AGP 8x or even PCI Express, this is incredibly slow transfer compared to using prebaked, preloaded, local-memory textures on a 512bit-wide 1GHz bus like in the latest graphics cards have)

The EFIKA board, like the PegasosPPC, implements Genesi's Open Firmware implementation which means the board boots and operates with a Forth interpreter on boot, has network and disk booting capability out of the box, supports USB keyboard, serial terminal, and/or VGA display with x86 BIOS initialisation and interrupt emulation for initial bring-up. Linux can query the OF for all it's startup information and you can run a generic PowerPC kernel (from Debian or Gentoo distribution CDs) without modification. As far as I know this is much better than any other embedded board you will buy on the market which run DINK or U-Boot and have a "special" board support package that comes with it with special kernels and patches.

It measures barely bigger than a 2.5" hard disk (and in fact has screw mounts on the board to attach a 2.5" disk directly to the board, over the RAM and processor). The entire thing without disk requires around 1W of power, and with a 2.5" notebook drive comes in under 2W idle and 3W while churning the disk heads for read/write. With a 1.8" notebook drive it can be lower still.

Current cost projections for the EFIKA put it at $299 for the initial production runs, and lowering down as we make thousands upon thousands, to around $150.

An exercise for the reader: sit down and simply IMAGINE the potential of a board like this. In the next couple posts I am gonna imagine for you, so.. don't think too hard.

Some pretty pictures of the board can be found in a gallery on the product website.

The End

2005-08-09T20:48:00.000-05:00

James didn't make it.

Update on James

2005-08-01T22:14:00.000-05:00

After some more complications with bleeding and so on, James stabilised to some degree and was put on a machine.. but he has to have a heart transplant. He is being further stabilised (he is getting better by the day despite being in a very critical condition) and then transported by special helicopter to a hospital in Little Rock, Arkansas.

There is hope, and these guys in Little Rock are supposed to be the best in the country. The transportation and then the transplant aren't without risk though. Friends are organising blood drives and donations.

Hang in there, James. You're in good hands. Keep fighting.. we're all behind you.

A slight diversion..

2005-07-29T22:30:00.000-05:00

I said in my first post to this blog that it would be about technology and business interests but I want to take a time-out from that.

A very good friend of mine, James, went for open heart surgery this morning to fix up a congenital defect (transpose heart) that he has had since birth. It's supposed to be his last surgery before he can start living a normal, energetic life. The surgery went fine, but on taking him off the machines, he is having serious complications and his heart keeps stopping. They are doing tests now to work out what the situation is. We find out tomorrow after some tests if he is brain damaged from the problem after surgery. If he is too far gone they are just going to turn him off.

All my thoughts and hopes go out to him and hope he pulls through and is okay tomorrow. James doesn't deserve this and if there is any good in the world he will be fine. If worst comes to the worst he will be put at the very top of the heart transplant list and get a new heart.

James, the world would be a shitty place without you. Please don't die.

Granny Smith

2005-07-01T16:54:00.000-05:00

http://developer.apple.com/hardware/ve/sse.html

New page today describing SSE and how to port AltiVec. Again doesn't show SSE off in the most favourable light in my opinion. It's not like Apple need to convince anyone (they will switch regardless of developer or user opinion) but still..

MIT Weblog Survey

2005-06-27T19:36:00.000-05:00

Make some science today :)

Apples and Oranges

2005-06-27T09:59:00.000-05:00

I love Apple today. They have single-handedly written a document (for the purpose of helping developers transition between PowerPC and x86) which shows off all of the flaws and foibles in x86 architecture, and makes PowerPC and AltiVec shine with heavenly light.. (edit: these links don't work too well anymore.. but they point to a header page which links off to similar documents)

Differences in Instruction Architectures
Aligning Data
Intrinsics Equivalents

I knew x86 ISA and SSE were bad anyway but wow these documents drive it home better than anything I've seen before.

Update, 28th June: MacSTL Benchmarks for those who don't beleive :)

The 64-bit Divide

2005-06-20T21:21:00.000-05:00

The craze for faster processors has taken us to a strange place - one where the MHz myth has finally fallen on it's ass, and now something I remember from games console marketing has reared it's ugly head - 32-bit vs. 64-bit.

There is no doubt that AMD64/EM64T architecture enhancements improve the performance of the x86 processors it is implemented in. x86 has always been register starved, with 8 General Purpose Registers to use and 3 of them permanently allocated to the OS in most circumstances. Now with 64-bit compiled applications using another supplemental 8 GPRs for both integer and SSE units (AMD64 and EM64T seems to have eschewed a discrete FPU) there is a lot less thrashing to be had in applications. Some complex routines on the old model would spend 1 instruction of every 4 doing register management, or unschedulable instructions which target memory (a CISC trait). Now I suppose this has been removed at least until 1 instruction in 12 in the most complex register-intensive operations. Compilers have an easier job and so do application developers.

When people apply this to PowerPC - or any other 32-bit/64-bit architecture like MIPS or SPARC - this is where they fall down. There is no such magical advantage on these processors. The difference between 32-bit and 64-bit PowerPC is a few instructions for dealing explicity with 64-bit datatypes, 64-bit GPRs (r0-r31) and the ability to access a greater memory area by default. Even 32-bit PowerPC's however, such as the G4, have the ability to address a 36-bit physical address space and a 52-bit virtual address space. Construct your OS to use 64-bit pointers and you would not see the difference. There are no "extra" registers. The FPU works identically. AltiVec is still the same unintrusive powerhouse. So where is the performance improvement that warrants everyone pining for a G5?

Some might say that the extra memory space is useful but nobody has exploited this at all. Apple's own hardware has a maximum installed memory size of 2GB or 3GB depending on model. You look at x86 motherboards (try Asus' website) and AthlonFX and Opteron motherboards are specified with maximum memory sizes of 2GB and 3GB again. Until you look for high-end server applications (such as those Intel has traditionally targetted the Xeon at) there is very very little support for the 64-bit memory space, and where there is, it was already supported using 64-bit pointers, and extensions like PAE or PowerPC 36-bit addressing.

In fact 64-bit systems are traditionally SLOWER than their 32-bit counterparts. If you look at the general operation of a 64-bit processor you can see that it takes twice as long to load in data to registers because they are double the size. For PowerPC, because of a fixed instruction word length, to load an "immediate" value into a register (that is a constant stored in source code, like 927464329 instead of "x") takes 7, 8 or 9 instructions where it could have been done in 3 previously.

Caches are usually doubled in size (data at least) and cache lines are also doubled. This increases latency, meaning that if the processor mistreads it has to do twice the work in order to fix the mistake. For faster memory buses needed to compensate for this, latency is increased further. For the G5 this is to the point that it takes DOUBLE the time of a G4 or a Pentium 4 to do a memory access. There are some benchmarks available from Freescale which show the G4's general prowess at doing random memory accesses like indirect accesses to arrays, where the G5 loves to do sequential streams. This changes the performance of software completely from a more generic performance of the G4 to something very specialised on the G5.

Instruction execution latencies are also increased - by 1 or 2 cycles each instruction sometimes. This means that clock for clock, the G5 will do LESS work than a G4. There is a point where the G5 is currently clocked just fast enough and ramped far enough that the performance levels are seperate (1.4Ghz and 1.8GHz respectively, you can see this in the eMac and iMac lines from Apple) and therefore you see a slight improvement rather than a drop, but check out Apple's new PowerBook - 1.67Ghz. That minimal 133MHz difference between the PowerBook and iMac means that when you are going to start noticing when latencies hit, when instructions don't execute as fast.

Why does Apple use 64-bit in the iMac, then? Well, it's a marketing game. They are playing the numbers, despite Steve Jobs almost coining the phrase "MHz Myth" he has been pushing for a 3GHz processor with 64-bit (magic numbers..). He even mentioned it in his keynote, saying that he promised 3GHz and didn't manage to produce it and that was one reason for the switch. You look across the pond to the market they are about to enter (64-bit enhanced x86) and see that AMD's top of the line processors barely top 2.6Ghz. Apple have test benchmarked against 2.4Ghz Athlon 64 processors for their website and come out tops with a 2.0GHz G5 processor.

Yes, we know Steve's reason for the switch is the APPARENT advances in Intel's processor technology, which may or may not include 30nm gates, 0.5V processors, and 10W power consumption at 2.5GHz. But sure even less than 2.5Ghz is the sweet spot; for dual core and multi-core you are better disposed to creating many slow processors to spread load than lots of very very fast very very hot very very inefficient resource-sharing ones. Sun's recent multi core offering was 1.2GHz per thread. That's a functional equivalent to a 2.4GHz Pentium 4 to you and me. A 2.5GHz dual core processor would be like a 5GHz Hyperthreading processor. Why bother to ramp the numbers up when you can simply multiply the number of real processors in the machine? His reasons are absolute crap and as and when we move to dual core in earnest, we will see that in actual fact Intel don't have the greatest technology in this field after all.

So. Don't tell me you need 64-bit :)

Betting The Orchard

2005-06-06T14:01:00.000-05:00

Oh me oh my! What have Apple done? They have announced 18 months in advance that they are switching to a new processor line.

Who would buy a Mac Mini now that a new version is due in 6 months? PowerBook Pentium in the next 10? Apple are betting themselves over such a long period in hardware sales I wonder how it will affect their profits? Now we will see how much their $20bn in the bank will last.

Hello, Tiger

2005-04-12T17:13:00.000-05:00

Apple's new OS and all it's neat features almost makes me want to get a Mac just to run it.

And then I am pulled back from the brink of putting my Mastercard number into the store, by the realisation that I really have nothing I want to run *ON* it. MacOS as an operating system is so excellent but I am at a loss for anything I personally need it for.

Is it healthy to lust for it, regardless? :)

Skin of Evil

2005-03-12T07:05:00.000-06:00

Oh how I loathe skins for windowing systems. How I loathe the pretentious design of 99% of the skins that are generated for them - their 8x8 pixel buttons with no spacing, everything is gradientised, black to charcoal grey, hyper-low-contrast with "shiny" highlights and metallic brush (the Apple, the) and reinventing system glyphs.

Call me nuts but I like the default skin for Windows XP. I get people complaining about that and MacOS being "kindergarten" skins, for children and women. Hello? Puberty is over, guys. Having colours on things does not make you a girl or gay, and what's wrong with either of those anyway? :)

User interface design is ERGONOMICS - the science of designing something that won't cause the user fatigue or discomfort, both physical and mental. It is not a dick size contest. It is not a way to see how close to black you can get, or to "maximize the space available for apps" by making the window gadgets BLANK and 4 pixels square. Apple were told they had been stupid for not labeling their buttons with images, but of course they did something much better; COLOURS.

The human brain is attuned to colours, and sometimes more importantly contrast, in very distinct ways; XP has a big red close button because it catches your eye like you'd hope a large predator would in the bush. Red is a danger colour, the colour of blood. Apple use the same idea for their Close "Jolly Rancher". The traffic light analogy is great too. As for glyphs on buttons, Apple put these on when you hover nearby - self-assurance for the user for when they get there. Windows, they are part of a Truetype font which comes with the system (mm, hinted, antialiasable, scalable) as are a great deal of the system gadgets in Office. Hand painted glyphs cannot be scaled to the user font size. Creating 100 bitmap variations is no job any designer should have to do in today's world either.

Sigh..

I guess the world will not change in favour of people with bad eyesight or poor motor function, as long as there are thousands of geeks out there who want their display to match their Denon amp. I say it needs to though.

Solaris is not the Devil

2005-01-29T08:22:00.000-06:00

Well the contraversy regarding Open Solaris is starting; although I can't see what for.

People are whining about patent clauses in the contract and other oddities they think "can be used to kill Linux". Personally, I think Linux has nothing to fear from Sun's opensourcing of the OpenSolaris operating system and it's tools like DTrace and ZFS. Sun isn't out there to destroy Linux via licensing clauses :)

It would be so easy to dismiss Sun and OpenSolaris out of hand, say it's a ploy, they are cheapening up, or going out of business, but it is really more subtle than that.

Sun are doing it because Solaris is kind of getting stodgy - in a world becoming more and more hip to the benefits of opensource products, the comparitively slow release cycle of Solaris, closed development of security fixes and improvements (released on Sun's timetable, and not the customers') and of course the restricted machines it can run on (SPARC architectures, x86 and AMD64, usually sold by Sun and nobody else) are not ideally suited to this kind of market.

Sun doesn't just want 1000s of free coders contributing to do their work for them, per se, but the community and the excitement that these 1000s of coders would bring to Solaris. That is what is missing from Solaris development; EXCITEMENT! DRIVE! The will to move forward and create something. The Linux Community has this in bucketloads, as does the BSD derivative community - pick one, Net, Free, Open, Dragonfly - in fact Dragonfly BSD is a perfect example of innovation and drive taking an operating system into different territories and technologies. Sun just wants a morsel of this to re-envigorate their operating system, and distribute their technology further afield.

Gentoo have stepped up and are supporting their meta-distribution technology on Solaris. It already runs on Solaris but OpenSolaris is a whole new goal; it means Gentoo for Solaris can be, like Linux, a full operating system of it's own instead of merely a package manager. This allows custom distributions to be built from the ground up, simply allowing OpenSolaris to scale more easily from one device to another.

Add to that the PowerPC port - Sun dropped this a long time ago, circa Solaris 2.x, and it was then only available in unreleased beta testing form - which is generating some more excitement and drive. With devices such as the Genesi Pegasos being donated to developers, this excitement and drive has a focus and I hope there will be some result in the near future which shows an operating system reborn on a superb architecture like PowerPC.

I say we let Sun tap in on the excitement, I think competition in the OS market is healthy. I'm looking forward to booting Solaris on my Pegasos, at least. Oh, and let's all try to support Sun and "shine" as developers and a community - not outright resent them - so that they seriously consider open sourcing more of their valuable products. Java, anyone?

A bystander's critique of the Mac Mini

2005-01-25T11:01:00.000-06:00

I guess I am the worst person in the world to be critical of Apple's new devices, since I work for a company that is technically in competition in that market space (PowerPC computing).

Nevertheless, I feel Apple aren't exactly on the ball here.

A simple look at the Mac Mini on Apple's website shows a cute, dimunutive machine - much nicer looking than a Shuttle or any ITX case, you might think, less bulky and easier to find desk space (or floorspace) for. What they don't show you so prominently is the entirely huge PSU brick. Some people say the Mac Mini is the modern embodiment of the Amiga A500, and they are right - it has a huge heavy power supply that just gets in your way. On a Shuttle case, you get a taller machine perhaps slightly longer on the desk, but the PSU is integrated. Which one wins in terms of total footprint?

Any which way, you are buying a tiny machine which just happens to have some quite large power bricks and chunks of plastic and metal in order to operate it. The size "benefit" is quickly thrown out of the window.

Having all the ports, and the power switch, on the back is also a mistake. I can imagine many people hammering USB plugs into the back of the case, fumbling and reaching around the back with little sense of orientation. How many times do you have to plug your USB stick into the power switch, thus killing your Mac, before you get pissed off or remember the USB is on the left hand side?

Apple are from a stable of switching on the system via the monitor, and having USB ports there too.

So, you think maybe you'll just buy an Apple monitor. Bzzt! Since moving to DVI, now you need an adapter box almost the size of the Mac Mini again to convert power, DVI and a USB hub connection into the monitor-native ADC. That size advantage is out of the window yet again. You can see this kind of "setup" if you click to see the Quicktime VR galleries of the Mac Mini. It's a ludicrous idea. Clicking the link for 3rd-party peripherals shows a Bluetooth Logitech keyboard and mouse and a decent monitor, with a lot less cabling and fuss.

The system is designed to work with any off-the-shelf keyboard or monitor (BYODKM), granted, but are Apple CONSCIOUSLY trying to dissuade people from buying Apple peripherals?

While I'm on the subject of peripherals, Apple really do seem to employ a rather naive view of accessorising their products. If you take a look at the iPod Shuffle - a great little device for a great price - you see it comes with a lanyard for hanging around your neck. The biggest show accessories are the "armband" for runners and the sports case which is neon orange to really shout out to the rest of the world - and a cap-replacing battery expander (with neck chain again).

Why not just hang $200 around your neck, and buy a neon sign for your hat, "easy pickings"? Pickpockets carry around scissors and knives already for the mobile phone lanyards that are so popular, the iPod is just walking into the trap.

Here in the UK and I'm told around Europe there is a huge public drive by police and security companies to really get into a spirit of hiding your electronics. Pickpockets and theives love it when you parade around your Christmas Presents for them to pick and choose - and when they are attached to your neck or arm by a flimsy detachable cap, all the better! I don't think Apple thought of the real-world consequences of flaunting the Apple brand and style in public like that.

It's just a failing of some design process that "cool" is flaunted over real functionality.

The cool styling of the slot-loading drive; it gives a clean look but slot-loading drives are provably unreliable in burning discs, compared to tray loaders, it precludes the possibility of using "mini DVD" or "business card CD" media, and also provides no emergency eject mechanism in case something gets stuck. Is this really user-friendly? It looks so nice, but the functionality of the device is artificially limited by some "cool" design choice. This is a damn shame.. a damn damn shame.

As for the technical aspect of the Mac Mini - processor, using a real DIMM socket, wealth of peripheral ports (Bluetooth etc.), I really cannot complain. The thing is a masterwork of miniaturisation, and does offer what most people need at least in terms of connections.

The G4 is an insanely good processor. You look at the PC world and the competing low-power-consumption and fanless processors are of the ilk of the Via Eden and Nehemiah, Transmeta Efficeon, and AMD's Geode NX. Rather than being a "cut-down" processor like the x86 competition, the G4 is not in any way a compromise - it has simple design compared to the "performance boosting" things like complex instruction scheduling or high bandwidth memory links present in the Geode and so on, but that belies it's power which simply outclasses the x86 designs.

Simple benchmarks you can run show that a 1GHz G4 is capable of outperforming a 1.4GHz competitor by some 200%. Then the G4 has the best DSP engine in the world - AltiVec, which Apple laud as much as they can - which increases the power yet again. The Geode by comparison doesn't include vector operations, and the Efficeon and Nehemiah merge the FPU and vector operations into the same physical execution unit. This is a performance hit that even special compiler scheduling and fancy pipelines in the processors simply cannot work around. AltiVec operates as 4 seperate vector instruction units. The FPU works in parallel. The Integer instruction pipeline is seperate. All this power in such a tiny box. Mac Mini wins in a way that the ITX platform simply doesn't.

$499 for the Mac Mini is really value for money for what you get. I just wish they'd eschewed some of the classy design in order to add the VFM that is so obviously unexploited. Maybe for the Mac Mini 2?

Enigmatic first post..

2005-01-22T17:20:00.000-06:00

Just to get the blog started. It seems blogging is the cool thing to do lately. I used to have one on Livejournal but that's more a personal blogging sensation, this one I think will be a bit more technical and professional, related to my work and general things.

Ho hum anyway.