Slashdot Mirror

Well then - we're in the same ballpark. However, I'm not a pedantic asshole, err, bastard.

My friends call me asshole, feel free. I do think you are being pedantic though.

So, let's take you to school:

Always happy to learn.

Lets see...
"--- very fast devices at both ends. This generally means RAID 0, 5, 6 or some combination thereof like 50 or 60.
Raid 10 is more for redundancy in case a disk dies than performance"

Nope, no speed by sequence there.

Let's see - I dropped RAID 3,4,30, and 40. Guess what, they're generally not used anymore. (Note - for your pedanticness, in general, ie, normally, ie, not specific use cases,
they're not used anymore. I just hit someone's comma's in a sentence limit.) This is just a natural sequencing of numbers there.

3's are still in use, but they are for people stuck in particular hardware. They suck.

"Second, RAID 5 with 8 disks is certainly faster than RAID 0 or RAID 10 with less than 6 equivalent disks"

If this is true of your setup, then there is something deeply wrong with your controller setup.
Either you have a controller with a fast cpu and lots of cache + really really poor disks, or it doesn't really do raid 10.

Why yes, I do have controllers with buttloads of cache AND a fast CPU. Why would I run enterprise systems on the cheap?

You missed the point. for this to be true, your disks have to be truly awfull, way way out of scale to the controller.

"I have owned about 8 different types, although none from the last 2-3 years - ie, IDE/SATA)"
Ok, now I'm getting the picture, you are a Hobbyist.

Not exactly, I dropped out of enterprise system hardware support about 5 years ago. Since then, I checked out the initial set of IDE RAID cards, and discovered they still
didn't overcome issue #1 with IDE - namely, multiple I/O. This means IDE of any flavor sucks for performance. They're great as mass store devices though.

Getging better all the time.

http://www.redbooks.ibm.com/Redbooks.nsf/RedbookAb stracts/tips0458.html?Open

Lots of enterprise class people are running fibre attached sata nowadays. Some business just can't see paying $1500 more for a drive to get that extra 15%

"RAID 6 would require a minimum of 5 drives"

Sorry, its only 4

That's true. Raid 5 only requires 3. Raid 5 with 3 drives, or raid 6 with 4, will suck eggs and isn't recommended. It's kinda like running windows with a minimum configuration -
it'll work, but it will suck. Even with 4 and 5 drives respectively, it won't be great.

Granted.
Raid 5 sucks for writes, is passable for reads, and career ending on the rebuild. And if it looses 2 drives, it's garbage
Over and over I've seen it. One drive fails and the hot spare rebuild kills another drive, leaving you praying you beat the 60% restore failure odds.

Raid 10 has no parity to rebuld. The hot spare rebuild is just a re-mirror, and only affects ONE other drive in the array, plus, you can loose up to half the drives at a time.
Just having the quick rebuild without significant storage performance degradation is worth the price of the other drives.
With raid 10, you find out you lost a drive when the log reports send you email.
With raid 5 you find out when the pager and phone go off, and your explaining to the VP that the system w

Different locales do things differently, I imagine the submitter is European.

"I cannot think of any other OS that even approaches the maturity of Linux at this point."

IBM might just decide to knock the dust off OS/2. In many ways Linux and Windows and yes even OSX still have not caught up with where it was a decade ago. I don't know of a 64 bit kernel for it or its offspring eCS but it might be hiding in one of the IBM labs in Boca Raton.

Actually the advances in hardware, emulators and virtualizers are making real time simulation of an entire hardware platform API or a specific OS API more practical. I suspect that any posix compliant OS with well written emulators or virtulizers will soon make reduce a specific OS like Windows to application level importance anyway.

I realize that most here are aware of these efforts, but are some urls anyway for the few that are not.

http://www.xensource.com/
http://www.parallels.com/
http://www-03.ibm.com/servers/eserver/about/virtua lization/about/systems.html
http://www.thefreecountry.com/emulators/pc.shtml
http://www.thefreecountry.com/emulators/macintosh. shtml
http://www.winehq.com/

Matthew

In case you have not noticed, the CMOS scaling of Moore is undergoing some squirming and shrinking problems. Have you noticed the clock scaling of microprocessos has been a bit more stable in the last year or two?

Latest and greatest feature size on the transistor (now starting production) is 45 nanometers. Research devices are at 10nm to 25nm.

A useful collection of info: http://www.research.ibm.com/journal/rd46-23.html

I have been on the R+D floor of Cray (Chippewa Falls Wisconsin anyone?) and the systems used for thermal cooling of these machines is impressive. When they talk "need to limit power consumption" the motivation is that the power generating stations are limited to 10-12 Megawatts.

The comment on "the big machines that the federal government has are not listed" is probably correct (I Don't know for certain either...) but then 99.95% of the engineers inside Cray don't know what their computers get used for.

Go figure...

Not every machine carries IDE/SATA. To enough people, a cheap short term solution will never be the case. U320 can do the job, and it may have to the way things work with IDE->SCSI conversions only able to match speed for these initial drives.

I think there is a very important point that you are missing in your argument. IBM DB2 is now the only database platform that has been enabled for Ruby on Rails by the database vendor. DB2 is not relying on drivers/adapters to be provided by someone in the community. As good as some of the community-provided enablement can be it is no match to what a database vendor would provide.

Another important point. New technologies like Ruby on Rails are typically rejected by the IT managers until they see evidence of a major IT provider committing to the technology. IBM release of enablement for DB2 on Rails http://alphaworks.ibm.com/tech/DB2onRails will go a long way in bringing Ruby on Rails in the enterprise ... a heck of a long way further then "just another database".

As for Ruby on Rails not being "enterprise ready" ... isn't the same thing that just a few years ago people were saying about Linux and Java? No doubt Ruby on Rails still has some growing up to do but it is ready for a lot of applications in the enterprise today. The question is "is the enterprise ready for Ruby on Rails?". Having IBM put DB2 on Rails and posting these articles on developerDomain can go a long way to getting enterprises there. For my money, everybody should try to do a small project now in RoR if for no other reason then to see what their programming job is going to look like in a couple of years :-)

You should take a look at Starter Toolkit for DB2 on Rails. It is a complete package of everything that you will need including a copy of DB2 Express - C that is absolutely free to distribute as part of your finished application and it does not limit the size of your data like the other database servers do. It is a complete one-click installer for everything. The integrated installer is for Windows at this point but the one for Linux is in the works. Bookmark http://db2onrails.com/ if you are interested in watching this very fast evolving space. As for the IDEs, there are quite a few really good ones out there RadRails that just got an award as the best Open Source Eclipse-based Tool. Matt Kent, Marc Baumbach, and Kyle Shank have done an amazing job on the tool and a nice presentation at the RailsConf 2006 in Chicago. There is also RDT. There is even a set of add-ins called Saphhire in Steel for Microsoft Visual Studio 2005. I can just see all the sladhdotters rushing to get this one :-) and flame to death in the process.

Personally, I think IBM should be commended. Their developerworks articles are usually of good quality, easy to find, well thought out, timely, and free. When I'm searching for some tech topic (specifically implementation instructions/tutorials) with Google if anything from IBM shows up, that's the first place I'll go. Just the other day I used their ssh keychain piece when I was working on setting that up.

I just wish more companies would follow there lead and provide something of value.

Ok, having reviewed the following document, IBM Redbook BlueGene/L: Application Development I'll lighten my "it's no easy task to bring apps to [this] platform" statement. It does appear that IBM has done some considerable work on the APIs and MPI support. Like any big beasty, there's always something you have to do to code before it will run well. I'd say from having another look, that it's probably no harder to bring apps to BlueGene than it is to bring them to System X running Mac OS X. It ain't a picnic, but it's certainly doable for a lot of community and homegrown code bases.

Not quite. The front end run and service node run Suse, but the compute nodes run a CNK (compute node kernel) that supports a subset of the system calls that Linux does and can only run one process per core (i.e. two per compute node). That means your code can't fork().

Lots of good reading material here: http://www.research.ibm.com/journal/rd49-23.html

Specs here and yes, Suse

Do you think an Osprey would fit in this or this?

Add some external RAID and you should be good to go.

Do you think an Osprey would fit in this or this?

Add some external RAID and you should be good to go.

http://www-128.ibm.com/developerworks/db2/library/ techarticle/dm-0406bhogal/

DB2 Express-C is an great option too. Furthermore, it is not limited to 4GB of data but 4GB of RAM.

Extracted from the IBM DB2 Express-C product page:

http://www-306.ibm.com/software/data/db2/udb/editi on-expressc.html
----
DB2 Express-C, a version of DB2 Universal Database Express Edition (DB2 Express) for the community, is a no-charge data server for use in development and deployment.

Maximum processors: 2
Maximum addressable memory: 4GB

Business partners may choose to register for free redistribution of DB2 Express-C with their products and applications.
---

Considering that its a fairly painless process to become an IBM business parter, this should help a lot of people who want to use it with their own products.

DISCLOSURE: I work for IBM, but I copied this text from the product page linked above.

Hit F4

No no, for a SAFER IE experience, hit Alt+F4.

Seriously, this interview was an example of "title inflation". The guy's not the "Lead Project Manager" - (how can you even have more than one lead) he's :

At Microsoft, I'm one of several Lead Program Managers on the IE team. My team and I are responsible for handling all of the incoming customer & security requests. I also do a lot of security outreach and enjoy spending time at various security conferences worldwide.

In other words, he's not even a project manager - he's works with the guys who takes all the complaints from people (from the helldesk/helpdesk), organizes them, and gives them to someone else ... but giving him a title of "Lead Project Manager" sounds better. Just how many "Lead Project Managers" are working on the IE7 project, anyway?

Choice clueless quote:

The first lesson was that the Internet isn't an innocent place any more. When IE6 was under development 6 years ago, viruses were inconveniences and true Internet crime wasn't a concern

Viruses have been a problem with Windows for how long? Oh, right - they've been with us since the DOS days ... http://www.research.ibm.com/antivirus/timeline.htm

And Word macro viruses have been around since 1995 - the same time Windows 95 was released. By 2000, there were over 500 new viruses a MONTH ... and the easiest way to spread them was by the Internet. 15 new viruses a day was only an inconvenience if you were one of the lucky/smart people who weren't running Windows.

Well, at least its not like it really matters, since the *real* project lead probably is more clued-in (or at least you can hope).

I can understand your concern. However, after IBM backs this up, it forces me to do more research (which, I haven't finsihed yet obviously).

You can read more about the History of Lotus Notes, and see a picture of the 1996 Lotus Notes 4 release here:
http://www-128.ibm.com/developerworks/lotus/librar y/ls-NDHistory/index.html
The next release was Lotus Notes 5 and shipped in 1999. That's an ancient version too, and is no longer supported by IBM. Yup, Ozzie has been gone a long, long time.

Two four story cooling plants? Huh? Assuming about two football fields (115,200 sf) and 300 w/sf(1), their cooling load should only run about... 10,000 tons. Assuming they throw in good 1200-2000 ton centrifs, thats fewer than 8 chillers, each with a footprint of about 200 sf (including tube pull service space). Double it for pumps and random aux equipment, and it still doesn't add up to a 4 story building (one of them must be the cooling towers).

That better be a 4 story cooling tower with a 4 story thermal storage tank charged by the cooling towers at night. If Google is honestly putting in 4 stories of electrical cooling equipment, they need to talk to someone stat - before dropping 6 MW onto Oregon's grid that could be avoided by cashing in on waterside economizing and a tank (using towers, assuming that thermal pollution of the Columbia is not an option). Hell, if you want to go cutting edge use the waste heat from the racks to run an adsorption chiller or something. OK, that's a bit silly, but theoretically possible (and despite the ghetto website - the translation from the manufacturer's native tounge isn't so hot - there are adsorption chillers installed in at least one critical facility I know of).

Comeon google, don't be evil. Go and design an actually efficient datacenter for once. Talk to the local boys at LBNL. They can help. If that thing ends up full of CRAC units with those hideously inefficient fans and no airside economizer (equivalent to opening the windows for the 70% of the time outside air is cool enough to condition the space)(with proper humidity lockouts, computers don't melt in filtered outside air you know), you are evil. EVIL!

1. And for that they better have a very tight hot aisle/cold aisle setup, or some very non-google'esqe custom/complex water cooling (coils on the racks would be doable, and possibly flexible enough for Google to go for, probably in a cheaper custom buy than commercial systems). I've never measured 100 w/sf (although close is becoming common), nevermind 300 w/sf, and have only designed for 150 w/sf with 300 w/sf expansion potential once; that involved a 4' raised floor for a supercomputer center, but it might be possible overhead with very good air management. So, 300 w/sf average over the entire space, not just one rack footprint, is a reasonable max load guesstimate.

We use sametime at work, and I find it much more effective than msn/icq/yahoo etc. There are a couple of extended client versions available offering more functionality than the original (though this functionality is intended to be included in Sametime 7.5)
Notes Buddy and
IBM Community Tools (ICT)

Quite a reliable system, I've found!

We use sametime at work, and I find it much more effective than msn/icq/yahoo etc. There are a couple of extended client versions available offering more functionality than the original (though this functionality is intended to be included in Sametime 7.5)
Notes Buddy and
IBM Community Tools (ICT)

Quite a reliable system, I've found!

FYI, we do know the answer to that question.

Get involved with your LUG. If you get involved you may find they are involved in many projects in your community. They may be involved in FOSS projects, network installations ... They will have an email list where you can ask any question you and should give you near instant responce. The LUG in my area is run by the Univerity professors that teach linux.
IBM has online stuff on getting your LPI.
http://www-128.ibm.com/developerworks/linux/lpi/
I think you can take the exams online for free as well.
Frozentech's Live CD site is a great place to dl distros and give them a test. You may want to think about what you want to do with the distro (ie clustering) and then find a distro to test and see if it fits your needs.

• Sun Microsystems Solaris and Linux Training
• IBM AIX Training
• IBM Linux Training
• HP's HP-UX Certification Training
• HP's Tru64 UNIX® Training
• Red Hat Training
• Novell SUSE Linux Training
• HP's NonStop UX Training
• Apple's Mac OS X Server Training
• And, if you're really sick... SCO's SCO UNIX Training

• Sun Microsystems Solaris and Linux Training
• IBM AIX Training
• IBM Linux Training
• HP's HP-UX Certification Training
• HP's Tru64 UNIX® Training
• Red Hat Training
• Novell SUSE Linux Training
• HP's NonStop UX Training
• Apple's Mac OS X Server Training
• And, if you're really sick... SCO's SCO UNIX Training

HP or IBM?

Personally, IBM research and development puts me in a constant state of awe. I believe they have some of the most brilliant minds in the world pushing the boundries of science. Maybe thier end products don't always reflect the level of R&D invested, but don't kid yourself... the last thing HP wants is IBM's full, undivided attention at it's market share.

IBM's strength is in it's diversity. Just because they cut PC's to Lenovo doesn't mean anything about the future of the companies presence in the future technology market.

Remember this little gem?..... http://www.research.ibm.com/quantuminfo/teleportat ion/index.html

The TIS Committee grants you a non-exclusive, worldwide, royalty-free license to use the information disclosed in this Specification to make your software TIS-compliant; no other license, express or implied, is granted or intended hereby.

Well, IBOC sounds good until you look at the spectrum data (from Ibiquity) http://www-128.ibm.com/developerworks/wireless/lib rary/wi-roam46/?ca=dgr-lnxw01HD-Radio#figure2

It clearly shows that with HD Radio, AM transmission consumes 30 Khz of bandwidth, which is (as I count it) 3 AM channels (since each AM channel is 10Khz in the US).

BTW, as you can see in the ibm article, FM IBOC also eats into adjacent channels.

Are we talking parent-child hierarchy tables? If so, Oracle's had statements to take care of that for a long time, since 1998 or so. Perhaps not ANSI standard, but they get the job done.

No, I'm talking about Common Table Expressions (okay, so I was slightly wrong about implementation of CTEs -- apparently other products have implemented the standard, but DB2 and SQL Server 2005 are the only "Big Boy" engines with them). CTEs aren't so much about implementing a hierarchy as they are about doing recursive actions quickly and efficiently. Walking a parent-child hierarchy is just an example of a recursive problem that's easily solved with a CTE, but CTEs don't dictate how you should store your relationship information.

For what it's worth, it's possible to write recursive algorithms in just about any SQL implementation (convert your recursion to iteration, and it's not so bad), but the win with using CTEs is that it's still a set operation. Doing the loop yourself means you're losing SQL's set-based power. I did a little comparison on a naive parent-child implementation, doing two things: return the path to parent from a given node, and return the subtree of a given node. I implemented each algorithm in SQL 2000's T-SQL without CTEs and in SQL 2005 with CTEs. The CTE implementation was approximately 10 times faster than the by-hand iteration solution.

http://www-106.ibm.com/developerworks/i/p-loeb.jpg

My company (a big one - on NASDAQ since 1983) uses a bit customized Rational ClearCase (Solaris) to store/manage all products and documents:
http://www-306.ibm.com/software/awdtools/clearcase /
Documents are written in SGML/XML (with DTD validation) - with help of customized Arbortext Epic editor:
http://www.arbortext.com/html/epic_editor_overview .html
It works pretty well!

/Z

Either that, or a broken benchmark. Each Cell processor (Synergistic Processing Element -- SPE) shares its instruction fetch port with its data memory port. The SPE can buffer up 80 instructions at a time (2.5 fetch words), plus an additional 32 from a branch target. Fetch will stall if the memory system gets saturated with loads and stores. Properly written memory-intensive code includes explicit fetches to keep these buffers full. Incorrectly written code will cause problems. Still, that doesn't explain a 3 orders of magnitude drop.

If you look at the slides on the page I linked to above, you'll see the SPEs are not connected into the global address space. They connect to a private single ported memory, and to each other through two unidirectional rings. (The ring structure is not apparent from that diagram, but trust me, it's there.) These rings then connect to a DMA engine.

If you wade through this paper, you'll see that the Cell compiler implements a software cache. (The same paper also explains the instruction fetch mechanism mentioned above, BTW.) That is, it emulates a cache in software, using the DMA to actually move memory around. Depending on the nature of the benchmark and how it was written, it could be that the read benchmark spends all its time allocating stuff into this cache and waiting for it to arrive. Writes would be faster because the cache can "write behind" without having to wait for the allocation to happen, if the compiler is smart enough to know that the previous data will be entirely overwritten. So, if the benchmark goofed, then the results are meaningless.

Fact of the matter is that the SPEs are capable of reading 128 bits a cycle each (128 bytes / cycle across the 8 SPEs). Other benchmarks, such as the article recently posted to Slashdot about using Cell for scientific computation confirm that this thing hauls--and these are bandwidth-intensive tasks. The quoted paper did run some numbers on real silicon and showed numbers similar to their simulation results.

With all this in mind, I find it hard to believe that Cell is broken.

Either that, or a broken benchmark. Each Cell processor (Synergistic Processing Element -- SPE) shares its instruction fetch port with its data memory port. The SPE can buffer up 80 instructions at a time (2.5 fetch words), plus an additional 32 from a branch target. Fetch will stall if the memory system gets saturated with loads and stores. Properly written memory-intensive code includes explicit fetches to keep these buffers full. Incorrectly written code will cause problems. Still, that doesn't explain a 3 orders of magnitude drop.

If you look at the slides on the page I linked to above, you'll see the SPEs are not connected into the global address space. They connect to a private single ported memory, and to each other through two unidirectional rings. (The ring structure is not apparent from that diagram, but trust me, it's there.) These rings then connect to a DMA engine.

If you wade through this paper, you'll see that the Cell compiler implements a software cache. (The same paper also explains the instruction fetch mechanism mentioned above, BTW.) That is, it emulates a cache in software, using the DMA to actually move memory around. Depending on the nature of the benchmark and how it was written, it could be that the read benchmark spends all its time allocating stuff into this cache and waiting for it to arrive. Writes would be faster because the cache can "write behind" without having to wait for the allocation to happen, if the compiler is smart enough to know that the previous data will be entirely overwritten. So, if the benchmark goofed, then the results are meaningless.

Fact of the matter is that the SPEs are capable of reading 128 bits a cycle each (128 bytes / cycle across the 8 SPEs). Other benchmarks, such as the article recently posted to Slashdot about using Cell for scientific computation confirm that this thing hauls--and these are bandwidth-intensive tasks. The quoted paper did run some numbers on real silicon and showed numbers similar to their simulation results.

With all this in mind, I find it hard to believe that Cell is broken.

Also, if you have already agreed (would assume its a none disclosure type thing) then why are you breaking that agreement and posting the links to the raw documents?

Microprocessor Online has some an interesting analysis. Pay attention to page 8, where the PS2 "Emotion Engine" processor is compared to the PS3 Cell processor. This is an analyst report for the industry of microprocessors.

If you really want to dig into the details of the Cell processor, check out Sony's resources. You have to agree to a bunch of things to get to the pdfs but there's a lot of information in them. Another place you can find information is IBM's resource site which contains a lot of stuff including the programming handbook.

Microprocessor Online has some an interesting analysis. Pay attention to page 8, where the PS2 "Emotion Engine" processor is compared to the PS3 Cell processor. This is an analyst report for the industry of microprocessors.

If you really want to dig into the details of the Cell processor, check out Sony's resources. You have to agree to a bunch of things to get to the pdfs but there's a lot of information in them. Another place you can find information is IBM's resource site which contains a lot of stuff including the programming handbook.

I certainly am thinking twice about buying my Raven X60 notebook, which is just a Lenovo X60 Thinkpad renamed to 'Raven' and with Linux (ubuntu Dapper Drake in my case) pre-installed.

The ONLY reason I am not cancelling this order, is that I don't want to mess up the Emperor Linux folks who have already had to order the laptop on my behalf. They're good people and I don't want to jerk them around. However, if this had happened last week, I'd have saved myself a bundle and bought one of the cheaper 12.1" notebooks instead. When paying a premium for a luxury notebook, I don't want to be supporting a company (Lenovo) that has such a poor opinion of their customers choices. This will likely be my LAST Thinkpad purchase. Maybe I'll put a Tux sticker over the Thinkpad logo to hide my shame ;)

This turn of events is really surprising to me because, I thought that, part of the reason that IBM, sold off this division was because there was a conflict of interest with their Linux software consulting and the pressure that they had from Microsoft. I thought that a major part of the issue was that Lenovo would be immune to this pressure and be working in a country where the local consumers were more likely to purchase linux boxen than elsewhere because of the strong Linux push in China. Of course, this could be looked at another way by realizing that, while IBM couldn't knuckle under to the pressure from Microsoft without losing face, Lenovo could.

So, where are the IBM linux consultants supposed to get their Linux laptops from now?

IBM never seriously pushed Linux on their desktop computers. There was one Linux ThinkPad model (cancelled several years ago, IIRC), and no desktops. Sure, you could pay their consultants to support Linux, but the PC group wasn't pushing it. It's pretty clear that IBM felt Linux was only useful for servers and workstations.

I disagree. IBM has certified Linux on their workstations, servers, and laptops for a long time now. Visit www.ibm.com/linux to see their Linux "portal". Under "Products and services - Hardware" click on "Linux on personal systems". You'll get a table of some laptop systems that are certified for Red Hat, SUSE, Novell, and TurboLinux.

The only thing I wish IBM had done differently was offer a series of ThinkPad models that came pre-installed with Linux. You can already get them pre-installed with Windows, but pre-installing them with one of the certified Linux distros would have made it much more visible. All that changed when IBM sold their desktop/laptop business to Lenovo, of course.

IBM/Lenovo makes the ideal keybaord, with both a built-in trackpad and a trackpoint. Same great tactile feel and convenience of integrated pointing devices as IBM ThinkPad keyboards. Two models are available, one with numeric keypad and one without. Here's a link to the one without:

http://www-131.ibm.com/webapp/wcs/stores/servlet/P roductDisplay?catalogId=-840&partNumber=31P9490&st oreId=10000001

That article you linked to... thanks for the laugh.

PS3 1% as powerful as the human brain?!, yeah, and my ass is rated at 4 teraflops.

That means, that a supercomputer rated at 200 teraflops (based on the PS3's claimed 2 Tflops) is equivalent to the human brain. So, when Blue Gene/L is complete, it'll be 1.8x as powerful as the human mind. Can you smell it yet?

At least one thing is for sure, Blue Gene/L (or the PS3) will never be able to run on cheetos and coffee.

Yang Yuanqing was rumoured to be grinning, while wispering "..excellent"

I am surprised you did not mention Tivoli Storage Manager. TSM just about rocks for these requirements and is ultra scalable. You can also set it up in such a way that the backups will simultaneously go to multiple tapes for the most critical data and that data can be periodically audited for readability etc.

http://www-306.ibm.com/software/tivoli/sw-atoz/ind exS.html

Last name I heard was System i5 They use POWER5 processors The wonky IBM term is Integrated Language Environment

Last name I heard was System i5 They use POWER5 processors The wonky IBM term is Integrated Language Environment

So I'm pathetic ... what the hell is help anyways? :-P Just as a followup ... wanted to say thanks ... got it working ... I was having the following documented issue

OK -- so this part of the thread seems to have become mired in uninformed babble. So... some facts:
1) The SPEs are optimized for streaming algorithms and for "synergistic" processing. Context switching is therefore not something worth discussing for the most part. For streaming algorithms, the assumption is that the code is small and the data is large (and DMA'ed in/out). For synergistic processing, the OS runs on the PowerPC PPE, which is perfectly capable of running any OS with good performance (like in all of those great big IBM servers), while the SPEs use run-to-completion on processing blocks of interest.

2) SPEs are not like SSE. SSE is a vector unit. The "equivalent" is the VMX/AltiVec unit on the PPE. (I use quotes because, as another poster said, we all know that AltiVec is superior.) SPEs run full threads while vector units are slaves to the main processing threads.

3) Cell was designed for broadband applications. Games are one such. The concern in such applications is balancing the processing power with the memory bandwidth... keeping the processors fed. In commodity processors, the memory bandwidth is frequently the bottleneck, fixed with ever-larger caches, OOO execution, branch prediction, deeper pipelines, and more tricks. These tricks cost silicon area = both power **and** cost. Instead, Cell adopted a very high speed memory interface. Yes, theoretical peak is theoretical, but several demos on Cell have demonstrated memory IO several times what you can get with a commodity processor.

4) The simulator is available to all at http://www.ibm.com/developerworks/power/cell. Run it yourself. According to the discussion forum there, the sim is cycle accurate for the SPEs. It is correct that memory interactions **may** reduce the performance from the sim since the sim does not simulate that accurately.

The more accurate way to think of Cell is to think of it as a PowerPC processor and 8 DSPs hooked together with very fast memory and using the same memory space. Now THAT is something.

The RAM is XDR. The IOIF (to talk to other Cells) connection is 2 FlexIO ports. The bus itself (called the EIB ) is something like 300 GB/s. I agree that peak is never achievable, but it should be possible to get around 18 GB/s or so.

Give me a moment, and I'm sure I'll think of one.

It's possible that this is the case, however IBM is actively working on compiler technology to abstract the complexity of an unshared memory architecture from developers whose goal isn't to squeeze the processor:

When compiling SPE code, the compiler identifies data references in system memory that have not been optimized by using explicit DMA transfers and inserts code to invoke the software-cache mechanism before each such reference.

So for developers who want performance, the architecture is ideal. 2 Megs of L1-speed memory, a 25 GB/s bus servicing 8 processors each with 128 128-bit registers. And for the rest, it's still a high-performance programmer-friendly development environment.

Your point is not going unnoticed by IBM.

When I see Big Brother-ish proposals like this I'm glad my employer is showing some decency and respect for privacy: http://www.ibm.com/news/us/en/2005/10/2005_10_11.h tml