CO2 Output and Total Cost of Ownership

According to a feature in HPCWire today by Edward J Lucente, the CO₂ output from datacentres is set to increase to about 6% of the US total over the next decade. The data in the article apparently comes from The Uptime Institute.

Although the figures quoted in the article are poor, making comparison between the US and countries such as Netherlands, Malaysia and Argentina; if the data is anywhere near being right then there is something to think seriously about.

I don't mean simply the impact on the globe, I mean the effectiveness of steps that we, the broad IT community, are taking to reduce it.

A substantial part of the rationale for multicore was to keep power consumption and thermal dissipation (understood in this context as being roughly the aircon necessary to keep the datacentre temperature within reasonable bounds) low (on a throughput per watt basis) while power (in the data crunching sense) rose. The target was to keep the growth curve static. Indeed companies such as Intel and AMD regularly show how this was achieved. If the conclusion of the report is true then the design win that we made has not had the effect that was predicted. Since the only way that compute requirements are going is up we must look into why that might be.

Of course, we  don't know the basis of the calculations that The Uptime Institute made, nor more importantly the assumptions that the datacentre managers who were - presumably - interviewed made. However assuming that the conclusion and methodology are valid then the conclusion has to be:

1. that multicore power saving doesn't work as it is supposed to;
2. that data centres / their needs are growing faster than can be compensated for by the decrease in power consumption;
3. that multicores aren't being used efficiently;

Of these (i) seems to be ruled out. It is more likely to be a combination of (ii) and (iii). In fact (ii),(iii) go hand-in-hand. If processing power isn't used efficiently datacentres will grow in order to meet demand. However if it is (see for example piece on Pervasive) then power consumption per item processed drops substantially.

Do we need yet another step change in technology? or should we put a lot more effort into using what we have better. We certainly haven't made best use of MCPs yet and the technology is planned to get much more sophisticated. HOWEVER if the figure quoted are correct then we have to do more as an industry to contribute to reducing CO₂ emissions.

Pervasive release

Pervasive (Austin, Tx) are a company that I have spoken of and written of many times before. Pervasive have just released a new suite for their Datarush software.

Datarush is very inteersting technology and represents somewhere close to best in the current class. Its JVM-based BI software demonstrates the potential for getting speed-ups of several hundred times on relatively limited numbers of cores. They have of course achieved such super-linear speed-ups through heavy re-architecting of applications and very careful coding, nonetheless they do demonstrate that large performance increases can be achieved that way. The software is available as library packages together with tools to build applications.

I don't quite agree with Robin Bloor that they are one of the companies to the top 10 companies to watch in the US IT sector. They certainly could be, but what they need at present is stronger leadership. If they had that, then Pervasive Datarush alone, ie the Datarush unit,  could certainly be one of the top US BI companies; the group would certainly be there. All that being said it is my present understanding that their clientele are very pleased with Datarush the product.

Is there life in FORTRAN?

FORTRAN isn't what it used to be. "Fortran" as we should now call it, is alive and kicking. But does it have a future?

I spent this afternoon at the BCS's (British Computer Society's) Fortran Specialist Group (http://www.fortran.bcs.org/), mainly because of multicore issues and the interesting variety of speakers. The big issue was - and remains - is the language still a viable tool? The answer from the group was a qualified "yes".

While Fortran has evolved overthe past few years its standardisation process has been painfully slow in comparison to the rate of processor evolution, and that of the market. Even today the 2003 standard is not fully implemented on most compilers. The 2008 standard will emerge later this year in very restricted form on NAg's compiler which would appear to be the only one (I am happy to be corrected on that). A next round is already underway with no clear idea as to when it will report, let alone be implemented.

One might ask what the value is of such a slow standardisation process. As usual the point is to enable portability and compatibility. Given the fact that Fortran codes are used the world over and in some pretty important applications there is an obvious need for portability and compatibility.

The problem in part has been that Fortran has been bringing itself into line with the structure and form of more modern languages (bear in mind FORTRAN was invented half a century ago). That was always bound to slow down things as fundamental issues were addressed and prejudices resolved. Today that process is largely sorted and the language looks and "feels" very different from its predecessors, at least in terms of capability. Now that the catharsis is out of the way perhaps future rounds can proceed with more speed.

But is there a market? The answer seems to be that there is but that it has changed. Fortran is alive, as a vehicle for implementing numerical models in industry as well as in HPC. This is true particularly in finance but also more broadly too. Vendors of compilers and other tools - and to a degree Concertant's own research - report a growing market.

However it seems to be becoming a language that people are picking up on the job and not being taught in colleges. Colleges seem to be teaching C/C++ and Java. Should Fortran be taught in college courses? Probably it should as a vehicle by which to learn scientific programming, for which it arguably does not have a peer, especially if you have a knowledge of maths. It is after all, FORmula TRANslation.

There is however an interesting discussion to be had about whether or not physicists  - and others - want to program using the language directly or whether they want to use a front end such as MATLAB or one of its alternatives. That is perhaps a discussion for another time.

Bull's Tera 100

One of the most interesting announcements that appeared at ISC last week was that by Bull of their Tera 100, 1.25 PFlop machine. At 4300 nodes and 140 000 cores which are Intel Xeon 7500 series-based it is no mean effort and claims to be the first Petaflop machine wholly conceived and designed in Europe. It was designed and built for France's CEA (Commissariat à l'Energie Atomique) - which has activities in four areas: low-carbon energy, defence and global security, information technology and health-related technologies - and its strategic goals. It is both a remarkable system and as a marker of Bull's comeback. In the late nineties Bull seemed somewhat of a spent force, kept afloat by strategic French needs (a view altogether at variance with then-prevailing Anglo-Saxon views of market forces). It had been a fairly important player in the seventies and eighties and at that stage represented - along with ICL and Siemens - a European force in mainstream IT and hardware. If you look back into the history of many of the European projects that fed into the mainstream Bull were a player at some stage in their genesis*.

Bull retrenched and concentrated on specialist projects and within the past few years have put several machines into the Top 500 each year from 2004 on. Today it has nine machines, with the highest machine at position 14. That is without Tera 100. If you include Tera 100 - with performance as quoted - then it would be among the top five. However all those machines are in Europe (in France, Germany and the UK). Whether Bull can be, or indeed wants to be, a long term player in the supercomputer market is another question altogether.

The next Top 500 will be very interesting for Bull as well as for others such as Dawning and for Cray and IBM!




*If you doubt this, every chip-and-pin card in your wallet owes its origins to work by a group of Bull employees who went on to form GemPlus which was the dominant force in smartcard technologies and made lead the way.

Knights Corner/Knights Ferry

It seems that I haven't blogged in some while. This week may well change things. ...


In any case, first to Intel and Knight's Corner/Knight's Ferry

Whatever the name and various people are using various names*, the KC processor, an implementation of what Intel are calling the MIC (Many Integrated Core) architecture looks to one more step down the many-core route for Intel and in-line with the aims of Intel;'s Terascale computing project. The Knights family is set to be a line of products according to Kirk Skaugen, vice president and general manager of Intel's Data Center Group.

Announced at ISC in Hamburg, Intel have now shown 80-core and 48-core processors before this; for neither of which they were prepared to give production dates at the time. Indeed they aren't doing so with KC either, nor are they saying exactly how many cores it will carry ("more than 50" according to sources). The processors will be implemented in 22 nm technology. Aubrey Isle is the codename of the silicon included the 'Knights Ferry' that forms the basis of Knight's Ferry (it isn't clear whether or not this is the same technology as that in Knight's Ferry - I suspect not)

 Intel Aubrey Isle processor die image

The MIC architecture itself shows a lot of similarities to the Larrabee architecture, so it appears clear where that has led. So not all of Larrabee has gone west. In fact it also owes quite a bit to Intel's 48 core Single-Chip Cloud Computer that was shown last year.

 Intel MIC (Many Integrated Core) architecture

In a real sense it is the convergence of the two. It is very much an HPC beast though at present. CERN Have been trialling one and they were quoted as saying that they were:

"... able to migrate a complex C++ parallel benchmark to the Intel MIC software development platform in just a few days...The familiar hardware programming model allowed us to get the software running much faster than expected."

Interestingly MIC is described as "An Intel Co-processor Architecture" with many, many more threads.

How long before we see the technology in "every home"? Well, not tomorrow, BUT bear in mind that performance levels even demanded by those who don't have to solve QCD or String Theory problems is rising inexorably. If we assume that Moore's law allows doubling at the usual rate, then assuming that we will see 12 or 16 cores this year then we are only two doublings off 50-core (or more :-) ) processors. That, it must be borne in mind is in theory at least, well within the reach of present technologies.

As a footnote, Intel are positioning many-core for the moment as follows: " Think multi-core technology is all about parallel processing? Think again. You can use multiple logical cores to consolidate platforms, provide longer life to legacy applications, reduce power consumption, increase overall system performance, combine a GUI and RTOS, and share system I/O resources". Not really best use, but...


* Strictly it would appear from the slides of the talk (http://www.intel.com/pressroom/archive/releases/20100531comp.htm) Ferry is the dev kit with "only" 32 cores and sits in something looking like a standard graphics coprocessor module.