Welcome to efficiency wonderland

Last post saw us pulling down average server instance consumption to the ‘Very Efficient’ less than 30 Watt server class. In order to get below this power envelope and within shooting distance of ‘Ultra Efficient’ (< 15 Watt) class servers we have to take the next step, this means throwing away the rule book and re thinking what a server is. So you have a choice; ‘You take the blue pill – the story ends, you wake up in your bed and believe whatever you want to believe. You take the red pill – you stay in Wonderland and I show you how deep the rabbit-hole goes’

Welcome to the world underneath the one you are familiar with, first stop down the tunnel is something that Via introduced called Nano ITX, it’s a motherboard (of sorts) with dimensions of a mere 12cm x 12cm, to put that into perspective it’s similar in size to cd/dvd envelope! So what do we get in such a format? here one of Via’s latest Nano ITX boards to wet your apetite,The VIA Epia-N700 Nano-ITX is a great example of this format and is one of several generations from VIA’s range.

VIA Epia-N700 Nano-ITX

If you checkout VIA’s description in the link you will gets some hints about the alternate world we are peering into here , take a look at VIA’s first line for the Epia-N700:

The VIA EPIA N700 is a compact, low heat, power-efficient Nano-ITX board, ideal for compact industrial PCs and embedded automation devices.

The key words here are ‘industrial’ and ‘embedded’ these words indicate “we’re not in Kansas anymore” this product is aimed at the embbeded/industrial market not the pc and server markets. The embedded market place is a different kind of environment, the hardware specifications focus on efficiency, tougher environmental ranges and integration with bespoke hardware and software. Most of the software development that takes place down here is C/C++ based with servings of assembly thrown in for optimum performance. Careful ‘here there be dragons’ this stuff is often used by real rocket scientists! But worry not, for our purposes and this post, we are going to stick around those environments we are most familiar with i.e. Linux and even Windows if you’re that way inclined. These boards will house those familiar operating systems allowing us to ignore some of the under-the-bonnet stuff. I will save such delights for the tinkerers later in 3fficient’s projects. Thus for now lets treat ths as a regular motherboard.

Now VIAs creation on this small format allowed them to bring X86 technology down to significantly reduced power consupmtion and board size to fit these solutions into much smaller devices than servers. We can leverage these benefits to produce mini servers that stack side by side on end at a density of approximately 12 servers within a 3U space, giving a 4 server per ‘U’ density. Along with that density comes a low power envelope in the 15 Watt region bringing ‘Ultra Efficient’ into shooting range. The only minor issue here is the performance the N700 comes with a choice of processors :

• 500MHz VIA Eden ULV processor
• 1.5GHz VIA C7 processor

The first provides a fanless solution, the second requires heatsink and fan. The choice of CPU will depend on server appliance workload requirements. The CPU combines with up to 2GB RAM (SODIMMM) and SATA connectivity for low powered SSD Storage. This product actually includes compact flash storage facilities onboard which can be handy but not for primary data storage due to performance limitations, except when your entire data set fits in RAM i.e. < 2GB. In most case one would add SATA SSDs to meet storage requirements.

However even though the N700 provides excellent capabilities in a small nano form factor, I was convinced that we could do better, particularly given Intels Atom announcements last year. Intel provide a chipset aimed at the embedded market called ‘Silverthorne’ part of the Atom family the two popular versions are Z510/Z530 both are single core atoms with reduced power envelopes. I am focusing mainly on the Z530 which clocks at 1.6Ghz and includes goodies like hyperthreading which a number of it’s partners don’t have. Actually Intel offer 5 versions Z500/510/520/530/540 clocking from 800Mhz up to 1.86Ghz with various TDP envelopes most around the 2 Watt mark, check out Ars Tehnica for a good overview and links to deeper pieces about Silverthorne. The key point here is that the Z530 and it’s combined chipset provide a very efficient and sturdy performance chipset made for this form factor, hence it didn’t take long to track down a candidate like Portwell’s NANO-8044 illustrated below:

Portwell's NANO-8044

This little Puppy takes the N700 performace much closer to the NC91 we discussed in the previous post Building rackmount hybrid clusters. Lets see what we get with this little beauty :-

• Atom Silverthorne Z510/530
• Up to 2GB DD2R RAM SODIMM
• Gigabit Ethernet
• IDE and Compact Flash
• PCI-Express

Not bad, the one big let down on this board is the lack of SATA although we could use a PCI-Express card to add this, or even use an IDE SSDs and save a few cents. More importantly this board operates within the ‘Ultra Efficient’ class i.e. consuming less than 15 Watts even under load conditions, excellent efficiency performance compared to where we started this campaign.

So lets review where we are, using these boards (or similar) we can construct high density (4 per Rack Unit) server arrays that consume less than 15 Watts per instance with relatively low enrgy loss in terms of heat, as these boards can be fanless (like the Nano-8044). In a hybrid cluster combining these with NC81s we can get the following cluster instance averages:

• 4:1 – Medium performance Efficient class 35 watts/webapp
• 8:1 – Low performance Efficient class 25 watts/webapp

That takes our average instance using higher ratios (6/8:1) on hybrid clusters well into the ‘Very Efficient’ class, not to be scoffed at! Additionally with the newer Opterons we may even be able to punch the 4:1 ratios into the same class for high performance clusters. In fact when I introduce some of the newer hybrid enhancements we are working on in the lab this target is easily achieved. For the moment however, I want to spend some time focusing on the larger efficiency opportunities that sit below the cluster hybrids that we have enumerated in the last two posts. I will be focusing on the longer tail of web apps that can run purely on bespoke solutions running within the ‘Ultra Efficient’ class.

In subsequent posts we will start getting into the subtleties of this new space and explore the new and innovative posibilities posed as the two great continents of PC server and embedded worlds collide, particularly with the emerging OpenSource influences transforming this edge market. As usual if you have any feedback, questions or suggestions please let me know or comment.

There is a great deal of innovation within arms reach in this area, it is ripe for energetic and passionate people to make a difference, I would love others to join in here, so if you’re interested twitter, email or just comment.

Building rackmount hybrid clusters

In the last post I talked about how power efficiency could be improved by moving from the popular vendor based 1U servers to the more compact dual ITX arrangements. The engines I recommended were based around Jetway NC81 ITX motherboards, which when equiped with RAID 10 made of 4 * SSD + 4Gb Ram providing plenty of grunt for webapps consuming 70-100 Watt. Given our ambitious targets even this level of power comsumption seemed higher than our original goals set out our Open & Efficient post for ‘Efficient’ class, clearly we have much more work to do!

So lets move forward and look at ways to improve the efficiency, to do this we need to understand our application space a little better. Most of the web apps we deploy (and like wise most others) are either 2 or 3 tier based web apps, with the majority of those being 2 tier. Tier 2 web apps have slowly overtaken the 3 tier apps over the last decade, with J2EE apps dominating the remaining 3 tier sector. If we take the 2 tier case as a model we find that there is a convenient way to divide the processing up. Tier 1 represents the front facing web server and business logic made up of controllers and views in modern MVC based web apps. Tier 2 represents the model and storage layer (although tier 1 often houses a thin model veneer or facade). This split of processing enables us to tackle modern web application stacks such as LAMP/LAMJ/LAMR or even j2EE allowing a hybrid approach. The hybrid requires a tier 2 transactional storage engine component and a tier 1 http request server and business logic. In most cases it’s the Tier 2 transactional and query layer that requires the greatest processing power and strong read/write subsystems. This contrasts to the tier 1 layer which is optimised for read (and cached) coded logic and process. Given this lets construct our hybrid datacenter recipe.

Tier 2 is easy to deal with as we have just the confguration in mind from our last post : our NC81′s provide just the kind of grunt we require for the transactional storage and query requests. But what of our tier 1? We require something more light weight and lower powered in order to meet our reduction energy goals. Luckily Last year Intel entered the market with a new range of power efficient processors under the moniker ‘Atom’. Whilst others were thinking low powered laptops, I was thinking power efficient servers that just shows my warped view of the world I’m afraid! Intel also did the honour of building ITX boards housing the new Atom wonders as can be seen below.

Intel D945gclf

How convenient is that? unfortunately there is a nasty gotcha, these ITX motherboards are not designed for rackmount server use, rather they sport a tall heatsink/fan combination adequate for desktop applications but not for low profile 1U servers. It is uncertain whether this was done on purpose or not, but the results is something that is unusable for our purposes. remember we want to integrate these motherboards into our existing datcenter architecture of dual ITX 1U units using the C147 rackmount cases.

Luckily our good friends at Jetway have come to the rescue with their own Atom boards designed with a low profile ITX format and the kind of quality you epect from these guys. Enter stage right Jetway’s NC92 series ITX inexpensive motherboards populated with either single core 270 or dual core 330 Atom processors.

Jetway NC91

Now obviously these boards are not in the same league as the NC81, but it turns out that they are actually quite nippy. For a tier 1 solution the NC91 populated with a dual core Atom 330, 2G DDR2 RAM, RAID 0 made of 2 x SSDs (hi-perf read ssd) offers great read and business logic performance more than adequate for a large range of web apps built around Java,Python,Perl,Ruby and PHP. remember that the store and query has been offloaded to the NC81s. The biggest win here is the power comsumption which is in the 30 Watt range.

Depending on application load the hybrids can be created with ratios of from 2-8:1, that is many NC91s to each NC81 in a cluster on a gigabit switch. Lets say we take an average config over 5U serving 8 high load web apps. We would have a single C147 containg 2 * NC81 (~160W), and 4 C147s containing 8 NC91s (~240W) consuming approximately 400 Watts or approximately 50 Watts per web app.

This hybrid approach enables us to slide our web application power envelopes into the ‘Efficient’ class (less than 60 Watts) efficiency band we set aside for Quad core performance. In this case we are actually working with the equivelant of 3 cores, two from the NC91 and a quarter of a NC81 (quad core). One can actually tinker with the ratios between tier 1 and tier 2 units to cope with different load patterns and requirements, below illustrates power bands for each ratio.

• 2:1 – Highest performance worst consumption 70 watts/webapp
• 3:1 – High performace ‘Efficient’ class 56 watts/webapp
• 4:1 – Medium performance ‘Efficient’ class 50 watts/webapp
• 8:1 – Low performance ‘Efficient’ class 40 watts/webapp

The last of these configurations would be used in situations where the web applications were known to have a lower requirement on the storage site, perhaps the database storage was only being used for metadata and the real data was in the form of images or other media stored on tier 1 rather than tier 2. Realistically 4:1 is likely the most frequently used configuration as it caters for most of the mid range apps at the top of the tail/bottom of the head curve. The power envelope for these hybrids is unlikely to drop as low as the next efficiency band (< 30 Watt), unless the consumption of the NC81 can be significantly reduced, this may be improved with the new series of AMD Opteron HE range or equivalent Quad core Phenoms when they arrive. This alone will not enable this solution to penetrate the next efficiency class barriers we have set our selves up with.

If we drop below the radar for the web applications spreading out before us in the long tail the performance requirements drop off even further and in many of these cases the top end of these could actually be handled completely by using NC91 configured with single core Atom 270 and run everthing including backend transaction storage and query on the same board. This would enable us to just break into the ‘Very Efficient’ class (< 30 Watts), but the victory would seem hollow given that the NC91 solution’s power consumptions is swamped by components outside of the CPU. Remember That a single core atom operates in the 4 Watt TDP range, a dual core in the 8 Watt TDP range, thus there is clearly a much bigger problem here in the architecture of this solution itself. For us to enter the holy grounds of ‘Very Efficient’ we going to have to look at more radical architectural approaches. Thus in the next post we will move on beyond ITX and even what we would normally view as motherboards in the server space and into a whole new world of wonderously efficient appliances and hybrids.

* Tip If you were considering deploying single core ITX boards, Jetway now offer the NF94 series which utilises Intel N270 (mobile version), this will shave off an extra watt or two and it all counts!

Simple off the shelf efficiency

I have been busy over the last few weeks and months shorlisting components and modules to build our prototype ultra low power servers. Prior to this however I also spent a lot of time considering low power servers using off the shelf motherboards as part of our previous rack mounted experiments. Although we decided to head for even more efficient servers the research could well be usefull for a class of data center applications, particularly hosting. Depending on configurations and chosen motherboards savings in the order of 30% can be acheived powerwise.

In order to improve our current hosting facilities we were choosing the following as a priority list compared to the offerings of multi-cpu/core servers like Dell’s PE 1950 and similar:

• Lower power
• Lower cost
• High density

Our goal was to build servers capable of handling busy 2/3 tier web applications and as such we were stacking them up against market favourites such as Dell’s PE 1950 and the like.

In order to get efficient motherboards we narrowed our choices down to AMD’s AM2 technology, this enabled us to take advantage of their excellent lower power range of processors in the dual and quad core variety. In order to meet the higher density requirements set for the project we opted for an ITX dimensioned (17cm x 17cm) motherboards. Initially we were attracted to Jetway’s NC62K Series but it’s single DIMM socket design seemed like a real compromise given the power capable of being placed in the CPU socket.

Jetway NC62k

Luckily Jetway had a newer motherboard up their sleeve (being sampled at the time) this was the NC81 series. This motherboard design used SODIMMs to enable fitting more than one dimm on the board, this allowed 4G of DDR2-800 to be installed as standard. The board would also sport a CPU with a TDP of 65W which took us into Quad core territory, awesome! Some of the other great features were 4 onboard SATA 2 ports with RAID 0,1 and 10 built into the bios in addition to 2 Gigabit ports.

Jetway NC81

This ITX format card has a real punch to it and Jetway produce high quality goods including good power designs and components (Advanced 3+1 PWM Design W / OC-CON Capacitors). As for power conservation you need to stick to the Phenom and Athlon ‘Energy Efficient’ versions of their ranges, examples below :

• AMD Phenom™ 9150e/9350e quad cores @ 65 Watt TDP
• AMD Athlon 64 X2 4850e/5050e dual core @ 45 Watt TDP
• AMD Athlon 64 LE-1640 single core @ 45 Watt TDP

As you can see from the specs their isn’t much point in going for the LE-1640, the move from 90nm to 65nm allowed AMD to produce dual core with the same power envelope as a dual core. Unfortunately this still hasn’t happened for the *quad core but given the power consumption elsewhere on this board one may as well go for the quad core, particularly if it is being aimed a high performance/load web apps or virtualization where you will get more bang for your buck.

Integrating these into an existing datacenter can be acheived simply using one of the rackmount cases from Travla the C147. This clever server chassis allows 2 ITX boards to fit in a single rackmount case and thus 1U rackspace, providing the high density required.

Travla C147

Storage wise we would go for Raid 10 formed of 4 of the newer high performance SATA SSDs from Intel, Samsung or Similar driven directly of the board. These will provide excellent read and write balance, and will actually surpass regular Raid hard drive configs in random reads. Given the power and physical size constraints the SSDs are pretty much mandatory, it also helps significantly on web apps with heavy read to write ratios.

Depending on choice of processor you’re likely to be in the 70-100 watt power consumption ranges per loaded instance. This compares to something like a Dell 1950 III, (Energy Smart) loaded up with 2 x Quad Core and raid 10 of 4 SAS drives, which will run loaded around 240 Watts and would deliver the equivalent of 2 instances. As long as you don’t require all of the power of a 1950 for a single instance (think virtualization here) then two of the NC81 units will probably provide you with as much web app grunt.

The power use in this configuration (2 x NC81) will likely add up to 150-180 Watts vs the PE’s 210-240 Watts, yielding an average saving of perhaps a third of the power consumption. Obviously your milage will vary depending on exact component choices and storage etc, but a clear benefit can be acheived. In the case where you are serving only 2 instances, virtualised on the PE 1950 and 2 physically by the NC81 pair, the I/O will be significantly better for the latter, this is down to virtualization’s extremely poor I/O performance. The PE 1950 does however have one significant edge, its ability to install 16GB RAM, in a number of cases particularly in single instances it becomes the clear winner.

Conclusion

With certain classes of web app, those that sit between the long tail of basic performance requirements (low traffic apps) and those mythic beasts with exceptional high performance (heavy traffic loads), i.e. the mid range 2/3 tier web apps are ideal candidates for this kind of setup. It’s a really useful configuration to have in our data center recipe kit for that web app sector. In reality however, we have bigger efficiency fish to fry – the plethora of servers within of the longer tail of web applications, these make up by far the greatest volume and hence greatest energy saving opportunity. My next post will show you how to start moving towards these goals fusing techniques shown here with a more hybrid approach.

I would love to here about any of your experiences and experiments with data center hardware and configuration, do you have any interesting recipes for power saving and or cost saving?

* Update : Although the AMD quad core cpus operate with a 75 Watt TDP we expect AMD to announce lower powered quad cores any day now, in fact they have already announced the Opteron HE series with 55 Watt ACP, we expect this to trickle down into the Phenom line at some point soon. Also note the move from TDP (peak) to ACP (average), not sure entirely how to compare the two exactly, but would expect this board to cope with a 55 watt ACP cpu and thus allow even greater power savings. The Opteron HE annoncement can be found on the AMD Blog with links to more useful data about the differences this provides, if it’s as good as they claim it can provide even better milage consumption wise. Also here is the video introducing the HE series

Some history and context for our pursuit

Although we aren’t a large player when it comes to hosting we do like to think we are innovative on behalf of our customers. Thus we regularly re-examine the value we provide and look to ways to improve what we offer in the datacenter. 

Originally our web servers were actually very efficient compared to todays crop of products. Those early experiences in 2001 of owning and managing datacenter hosted servers were based around the Cobalt RaQ models. These early low cost rackmounted servers were designed around MIPs Risk processors running at 150-250Mhz. The RaQ 2 for instance was rated at a mere 35 Watts, unlike the power sucking servers of today, much of that was due to the efficient MIPS motherboards of the Time at these lower clock frequencies. Despite this relative low power we could run up to 10 clients on each box with basic Apache http and CGI. At that time our solutions were engineered around perl and later PHP. When we decided to improve and unify code re-use we made the move to Java and servlets for our web application development. It was this move we that effectively determined that the RaQs would need to be superseded, partly because running Java on these systems was a maintenance nightmare as it wasn’t included with the supported package management and also performance was beginning to show these servers age. At the time Cobalt themselves had begun moving over to an AMD K6 based architecture. 

This was the ideal time to jump up to the much higher performance and subsequent power consumption servers of the pentium world. Although we tried a number of different servers we eventually standardised in Dell Poweredge 1650 (1U) and 2850 (2U) rackmounts, these have survived as standard components in our datacenters until now. Throughou that period however, we did experimented with alternatives server sources in the UK. We had a brief flirt with a UK organisation called DNUK for AMD based solutions, but often found Dell could always win on pricing with Intel Xeons so that avenue quickly closed for us. There was however one experiment that is worth discussing here as it seeded some of the ideas that we are drawing upon today.

We purchased some servers from a very small specialist supplier in the UK which were very different from anything else around at that time. But to explain what took us there I need to explain our requirements and though processes that led us in that direction.

The reason we had settled on Dell PE 1650 and 2850 models was not really about performance per se, rather it was about reliability. When you run servers remotely you need them to stay up, they are not within reach to sort out when the fail. Sure we have remote hands that can perform simple operations like systems restarts and hot disk/tape swapping, but thats about their limit, it also costs a lot to use the remote hands so there is an additional financial burden for unreliability. Therefore we set some basic redundancy features on every purchasing cycle for new servers. The two main features where first and most important, battery backed hardware RAID (we standardised on RAID 5 for perfomance reasons) with hot swap 10K/15K high MTBF SCSI drives and secondly, dual failover PSUs. Storage being mechanical and power supplies being high voltage/current where the most common parts to fail, particularly during brown outs. Of course these features don’t come cheap High performance high MTBF SCSI drives were (and still are) very expensive compared to the modern consumer crop of SATA/IDE products. Bullet proof raid hardware also added significant burden and pricing along with the dual redundant PSUs. Dell’s Poweredge 1650 and 2850 models provide all of these features with a range of processing power at relatively competitive prices vs IBM/HP et al. Never the less, competitive pricing was still a whole lot in the early days, I remember our first Dell order for a single 1650 and a 2850 costing in the order of 5-6K GBP, significant capital for a small business like ourselves.

It was this pricing and lack of competitive choices that led us to investigate alternatives and we were looking to be innovative, ready to experiment in order to solve the cost issues. We came accross this very small supplier (who were later bought out by a larger outfit) that built very small aluminium 1U servers that would accept a very large range of motherboards and storage, they were also willing to build to order in very small quantities. One of the key features of their servers is they offered a 1U server box which was very short (similar to the RaQs) which enables to to sit two into a 1U space back to back. This idea appealed to us and we experimented in using them in this high density model (two servers for every 1U) as it was ideal for the low end basic virtual web serving market that we had expanded into. These ‘Half U systems’ were fairly basic, constructed out of more consumer oriented drives (the motherboards were more specialised) and slightly lower wattage single PSU (less than 180 Watt). We could buy two of these and run them redundantly and still have change compared to the ‘grown up’ servers. Unfortunatelly to our dismay we found these units to be very susceptible to brown outs, a single brown out could take out several of these servers in one fell swoop, Ouch!! this was not amusing to either us or our customers, and was mostly due to the use of lower cost drives, but occasionally even the PSUs popped, which turned out to be poorly designed.

Our short history of servers:

1.   Cobalt RaQ2
2.  Dell poweredge 1650/2850
3.   Generic half depth high density

 

That was our last full experiment with alternative server suppliers, not because of failures, but rather down to our change of business model over that period. We decided to focus purely on managing the applications we were building for customers and pulled back from the generic web serving market (the main reason for datacenter expansion). It made a great deal of sense retrospectively as there were large players emerging driving down prices to fit the huge scales that they could muster. 

Now ironically we find ourselves full circle to some degree and when I look back at those Cobalt RaQs I envy the power envelope we used to operate in, our experiments over the last few years have taught us a much in the interim and our current exploration into low and ultra low power servers is turning into a fascinating journey as well as a challenging enterprise.

3fficient – Open & Efficient

Some of you may be wondering why I have started another blog, Folknology (my personal blog) covers a generic set of subjects mostly around the tech industry, programming etc.. 3fficient however is a different kind of animal, it stems from a frustrating experience I (or more precisely my business) has had over the last several months.

As part of my day job I (and my colleagues) look after our customer’s servers and or datacentre requirements along with the software we engineer for them. Recently I embarked on sourcing some server replacements for our existing co-hosted servers which were approaching end of life. At the same time I decided to investigate how we could lower our environmental impact, more particularly our carbon footprint.

Our biggest ongoing carbon cost is of course power, this is particularly so in datacenters compared to other buildings. It is not just the electricity unit rates but also the inclusion of cooling which often doubles effective carbon energy costs. Often Datacenters will also charge above standard electricity  unit rates to cover their redundancy infrastructure although this is tends to be a controversial subject.

Thus my first priority was to source lower power servers suitable for datacenter environments, this was also where I hit a few suprises. The server vendor we have historically relied upon (Dell) have a new range called Energy Smart servers. If you look at their supporting documentation they have a Performance Brief showing one of these servers in action compared to an IBM and HP equivalent server. Although they clearly show power savings over their competitors we are still talking about servers operating in the +200 Watt range and I consider these savings to be in an entirely different league to those of our expectations.

The deeper I looked at this the more questions I asked myself, as a result rather than just buying the new replacement models superceding the older units, I decided to entirely re-examine our needs as a business. We basically redefined our requirements from our existing operations. Here are some of the key facts we discovered

1. We were only using a fraction of our storage space on each server.
2. We were only using a fraction of the compute power on each server, with the exception of some infrequent batch work.
3. We did benefit from the performance and redundancy provided by RAID (lots of random reads, occasional disk failures)
4. We did not benefit from dual redundant PSUs (most issues resulted from datacenter power brown outs and faulty power racking)

From this we derived the following basic requirements

1. The storage requirements rarely exceeded 20Gb, with pruning and archiving much less still.
2. CPU performance estimated from many of existing web apps – Most of these apps were happily catered for running on Dell 1650 rackmounts with single 1.2Ghz CPUs with only occasional peaks.
3. There are lots of random file access on the web servers and databases so we would like to get around 100MB/sec reads with low latency and low failure. (we have had a number of disk failures).
4. PSU need to be very efficient and reliable we would also like to move into 3 new low energy performace bands. (We would also like to solve the brown outs I will cover that another time)

From this basic analysis we could prioritise our requirements and factor in energy efficiency.

The bands we have chosen cover the applications we expect to run, most of which we would like to squeeze into the lowest energy band.

Target Energy Bands:

1. Efficient – < 60 Watt – Quad core performance, 4G RAM, 200MB/sec storage
2. Very Efficient – < 30 Watt – dual core performance, 2G RAM, 150MB/sec storage
3. Ultra Efficient – < 15 Watt – single core performance, 1G RAM, 100MB/sec storage

As you can see storage performance is really important to us, capacity is a lot less so. Thats because most of our apps require relatively low storage capacities and we like to archive outside our datacenters for redundancy, thus less is most definitely more for us.

Some of the folks I have talked to have laughed at these energy targets, and for datacenter server applications and I can see why they are skeptical, the entire industry is based around a perfomance plus mentality and my ideas here do not seem to fit with their product plans. Never the less I am deadly serious about this, setting difficult targets means radically re-thinking the server, it may also lead to re-inventing the datacenter. Either way if you want to come along for the ride I will be sharing my research here, if you want to join in the fun I will also be detailing how you can also participate.

Being a very small operation we clearly can’t change the server world alone, that is why we intend to share the software and hardware that we develop in this mission so we can all benefit. Perhaps we can start the real efficient server revolution here one server at a time using the power of collaboration and  OpenSource!