Buyers' Guide
Xeon Buyer's Guide: Flexibility for the future
Published: 27 Nov 2008, 04:23pm
The future of the data centre can be mapped out in several directions. New developments in hardware and software are driving a shift towards flexibility, while some predict the "cloud" will spell the end of the old-style IT department. But can the choice of processor help keep abreast of these developments?
Hardware and software have a symbiotic relationship. The processor evolves to give best performance with existing software, while the software develops to support new applications and make the best of the underlying hardware.
If hardware or software makers try to second-guess the next steps it can lead to problems, but right now information flow is good, and the industry is aligned. The paths to the future are clear, and hardware and software are in step.
The major trends now are multi-core in hardware and virtualisation in software. The Intel Xeon processor 7400 series, codenamed Dunnington, only recently launched, but the company has demonstrated 80-core CPUs in the labs, so we can be fairly sure that multiple cores are the direction in which hardware is going.
Don't expect 80-core Xeons any time soon, however. The 80-core project does not use x86 processing cores. Instead it uses a simple core that essentially demonstrates what could be done with the current manufacturing technology. But the project will certainly influence future developments.
Multi-threaded software development
"It is a challenge for the software industry to keep up with the shift in underlying CPU architecture," explains Intel solutions architect William Crowe. "Multi-core CPU architecture has had a significant impact on how software needs to be designed if it is to take advantage of the great increase in performance that multi-core can unleash."
The key thing the software development community has had to get to grips with is moving from single threaded code to develop multi-threaded code. In other words, these multiple execution units in the hardware have a fundamental bearing on how efficient code is executed and, obviously, this has a direct bearing on application performance.
The majority of standard IT applications have undergone this transition to become “multi-threaded” over the last few years. Multi-threading simply implies that the code is developed specifically to run in multiple execution units, in parallel execution.
Some applications remain single-threaded but these tend to be instances where there is little return on the investment to re-engineer the code. In such instances CPU selection is important because the code will effectively only execute in one of the CPU cores, with the other cores sat idling. For this type of scenario it would be desirable minimise the core count - typically two cores/CPU is the minimum - and have higher frequency (GHz) to increase application performance.
This is very inefficient from a power consumption perspective and the overwhelming change within the software industry is to develop applications to full advantage of the multiple execution units available.
Typically, well written applications can unleash dramatic increases in performance - sometimes 10-times improvement can be seen - when compared to running on the single core machines of just a few years ago.
To make these changes to the software there are various development tools that can be employed to ensure optimisation for threading and tuning. Tuning alone can often enable up to 30 per cent performance improvements, just by using the tools correctly. Clearly it can be seen that higher performance code can deliver direct TCO benefits to the business.
Virtualisation
Virtualised operating systems place multiple applications and virtual machines on the same hardware. Multiple virtual machines place greater demands on the hardware, as applications have to share access to memory, networks and other resources. More I/O is needed, so newer chips have efficient technology to share resources.
All this feeds back into hardware as those chips fit better in blade servers, which pack processors ever more tightly in racks. And the power consumption per processor has to go down so those racks can work without overheating or blowing the server room power budget.
Even so, blade architectures have taken the power needs of a typical cabinet in a server rack up from a typical 6kW up to nearer 15kW. To go any further would require a change from air-cooling to more sophisticated cooling techniques such as water-cooling.
The modern server room has an abundance of processor power on tap and, with modern management software, it is available to be provisioned with just a few mouse clicks. In future the trend will be to have this provisioning fully automated so datacentres can react dynamically as the workload demand varies.
This vision of fully automated datacenters providing this level of business agility is where IT is heading. Today, however, the latest advances are around virtualisation. Until now virtualisation has been used as a mechanism to consolidate physical servers. Moving forward, the advances are around the ability to dynamically move these virtual machines across physical servers. This is enabled by such things as VMware's Vmotion, which lets virtual machines move from one server to another without missing a beat.
Future compatibility and investment protection
New developments in the Intel Xeon CPU architecture makes this even better: previously, the processors had to be the exact same model in order to guarantee compatibility with the Instruction Set Architecure, but a new feature called Intel FlexMigration (Intel VT FlexMigration) in Intel's Xeon lets virtual machines move onto future Xeon based servers. Essentially, this enables future compatibility, providing investment protection and greater flexibility when designing virtualised resource pools.
All this is leading to a shift towards "cloud computing", where a service provider can use those clicks to deliver processing power to remote customers - more quickly and efficiently than the customer can do for themselves. Large providers are looking to open this up even further, delivering their cloud services from whichever country in the world can provide the cheapest energy costs and infrastructure at any given point in time.
These big datacentres need to be able to respond intelligently to the processor demand - spinning processor cores up and down as required by the application workload. "This is a future technology we are going to start bringing out shortly," says Intel’s Crowe. "We can aim for a realistic level of 50 to 60 per cent utilisation. We're trying to get away from the situation where thousands of servers are at ten per cent utilisation."
Open Source
But what's going to run on these clouds? There's a healthy level of competition, with open source providers such as Red Hat bidding to outmanoeuvre Microsoft. The good news is that the rapid flow of information now means all processors are equally able to support whichever software is chosen.
Intel's partnership with Microsoft is well known but the company is also in the top three contributors to the open source community, working with players such as Citrix Xensource, Red Hat, Sun Microsystems and others. “We have approximately 3,000 software engineers and a huge proportion of these are contributing to open source projects," says Crowe.
These software engineers are also developing the latest tools to enable software tuning and optimisation to improve the development cycle and ensure software code can fully leverage the hardware acceleration features. “We work extensively with most of the major software vendors to ensure their development engineers are fully aware of these advances in silicon development,” says Crowe.
One place that needs to keep up with the latest developments is the Cornell Institute for Social and Economic Research (CISER), where scientists investigate issues such as disaster recovery and food safety. The Institute collaborates with other bodies round the world, and data sets can be up to 75Gbyte.
CISER consolidated eight physical servers down to four after positive results in a test using one of the Institute's own CPU-intensive workloads. Four-socket servers, with four-core Xeon 7300 processors, were most efficient - handling the load in an hour, compared with more than three hours on a legacy server.
“With the Quad-Core Intel Xeon processor 7300 series, we have at least four times the processing capacity of our previous servers in half the floor space,” says Janet Heslop, IT associate director at CISER. “Our researchers can get more work done, and we have space for future growth.”
Digg
Slashdot
Del.ici.ous
Stumble
Reddit
