Sun is promoting their Logical Domain technology, something available as a free open source equivalent. Is the Sun technology a good idea? Reading their white paper is worthwhile with or without buying their products. Read Beginners Guide to LDoms with the following points in mind.
Logical domains are called virtual private servers by everyone except Sun. When you look for alternatives to Sun's Logical Domains, search for virtual server, vps, or virtual private server.
The Sun paper talks about saving hardware by increasing utilisation through virtualisation. Hardware is cheap compared to the cost of losing a customer. Increasing the utilisation of a computer from 10 percent to 30 percent is useless when you have a ten fold or more variation in use due to minor changes in customer activity. Most companies should be planning to decrease their Web server utilisation from 30 percent to 10 percent to increase reserves for peak customer visits.
The lower cost Sun hardware is standard Intel equipment available from hundreds of suppliers. Only the higher end Sun hardware has anything special and that hardware is up against close competitors from IBM, HP, and others.
The generic approach to hardware, used by Google and many sites before Google, simply brings together many low cost generic computers to share the workload and each computer deliberately has a low utilisation to allow for the massive variation in Web site usage.
The expensive proprietary approach is to use special blade servers where you commit to a brand, buy their proprietary blade servers and the proprietary frames for the blade servers for a premium price, then hope you will save enough money somewhere else to cover the extra hardware cost. Usually the incentive to buy expensive high density blade servers is extremely high floor space cost. When your computer room is in premium office space in the centre of the city, look at blade servers or moving to a lower cost location.
Sun mentions hypervisors. Hypervisors are virtual machine systems, as invented by IBM over 30 years ago. The best known current hypervisor is VMware. Xen is the up and coming replacement, with a free open source entry level product and an enterprise level product for a reasonable price. Xen offers a special feature where it collaborates with operating systems to reduce the overheads of a Hypervisor.
Consider the free Xen software for sharing a server among a small number of test systems. You can share 4 gigabytes of memory across four virtual servers, which is enough to run a couple of test systems, a source control system, and a knowledge base for a few developers. For anything slightly larger, a generic server is easy to install and manage. Jump into clusters as soon as possible to gain the experience and to test the ability of your applications to work in clusters.
Virtual Private Servers
If you have several Web servers with different usage patterns, you can combine them on one physical server to share the spare capacity. Sharing capacity does not work when the Web servers all peak at the same time. You might have luck combining Web servers if each one supplies a different time zone that spreads their peak workloads across the day and night.
When you start your own Web site and need only a small fraction of a server, you might use a VPS from a web hosting company on the basis that it costs only 10 dollars per month and only a few people visit your site. One link from another site can bring a hundred visitors per hour and flood your VPS. The Web hosting company should automatically upgrade your Web site to have more resources and increase the cost from 10 to 15 dollars. That assumes they have the spare capacity of their server. you might find you suddenly have to switch to another Web hosting company and pay 40 dollars per month to get a reasonable response time on your Web site.
The difference in dollars is both the cost of the capacity for your Web site and the cost of leaving spare capacity on the server for expansion. Yes, VPSs are a good idea when you and everyone else expands slowly but in a rapidly expanding Web world, the cost of a large expansion capacity has to be paid by someone. You can soon read a point where a dedicated server is as cheap as paying a premium to use a server with a large spare capacity.
You can place several servers onto one physical server if they can share resources on a priority basis. you might place a Web server on the same physical server as an email server and give the Web server priority over email. When the Web server peaks, email slows down. A few minutes of email server slowdown is acceptable but several hours of stagnation, as I sometimes see in the corporate world, is unacceptable, resulting in staff using expensive telephone calls and faxes to replace unreceived email.
There are many other ways to share servers and most do not work. If you place a busy print server at a lower priority than a busy Web server then you under utilise your expensive printers and hold up contract printing to the point where you lose sales. Databases peak at the same time as the applications they server which makes them incompatible.
Death by a Thousand Milliseconds
Look at the average corporate server. They are often overloaded within a month of installation because they were undersized when purchased. Your staff are delayed by a few milliseconds every keystroke and it soon adds up to minutes per day. Multiple ten minutes per person by 600 staff and you are losing 100 hours per day, equivalent to 12 - 15 people or 2.5 percent.
After 6 months the server is delaying people by many milliseconds and no one will replace the server because the server is so expensive. Somebody orders a memory upgrade and the premium priced proprietary memory blows the hardware budget for a month.
Now switch to clusters of cheap generic computers. Not every application fits clusters without a redesign which means you should look at clusters way back before you buy an application. The cluster of generic computers can be upgraded with a new computer every month for less than the cost of that proprietary memory. You get more processing power and more memory for a low incremental price. You decide exactly how much delay is acceptable and upgrade before your staff hit the speed limit.
Some brands of blade servers can be used like clusters of generic computers if you want to save floor space but few are anywhere near the low price of generic computers.
The Final Upgrade
No matter how big the blade server cage, eventually you need something bigger and then you are up for the big upgrade where your proprietary hardware supplier charges you a fortune to move up to "the next level". What is the magic solution they offer? A cluster.
Yes, you have to implement the same cluster system you could have implemented a long time ago with cheaper generic hardware. No matter which way you go with hardware and software, design clustering in from the start so you are not locked into a dead end system where you are continually paying a premium to buy the latest proprietary hardware.
Every time you commit to a closed solution, you pay a premium up front, keep paying a premium every year to maintain the system then hit a super high cost at upgrade time. Even if their service is better and their initial price is attractive, look at the total cost up to and including the upgrade when you hit the limit of the current system.
A long time ago people promoted distributed computing as a way to lower your central computing cost. At the time, distributed computing was limited by the low power of the computers at the user end of the network. Today those computers have heaps of power but sales people are talking customers into moving the workload back to the overloaded central computers. Look for ways to make better use of all that computing power out on the desktop.
There are lots of new ways to move work out to the desktop without losing control at the central site. Give me a call for more details.
Look at the Sun white paper, look at the Sun hardware, Sun software, and alternatives. Think about all those customers and staff sitting around doing nothing while your over utilised servers fight over the reduced capacity. Look at increasing the capacity of all your servers instead of strangling sales.