Editor's Note: This is the first installment of our new weekly hardware feature. Hardware Today will cover the latest hardware trends and issues, as well as offer purchasing tips and explanations of various types of hardware. We welcome your feedback and story ideas, which can be sent directly to author Carl Weinschenk or to ServerWatch.
It's a transition that is not happening quickly, but the world of server computing is irrevocably shifting away from mainframes and the familiar rack-mounted servers toward blade servers. Blades are the new kids on the block, but rack-mounts, which in some respects can be viewed as their predecessors, are well entrenched. Blades fit into racks, and it's easily argued they are the next step in the evolution.
Mainframes, despite their longevity, are also here to stay. They are "still terrific for high-end transaction processing as well as large-scale database processing," says Tim Dougherty, the director of strategy for IBM's systems group. "There is a set of applications, especially in the batch world, that it continues to do a better job of than any of the others."
In the big picture, however, modern IT is pushing toward modularization, a factor that lends itself neatly to the concept of blades. Demand for server power can be fulfilled in one of two ways: by "scaling up" or "scaling out." Scaling up refers to adding more processors and processing power to a single server, while scaling out refers to linking more units together, usually in a clustered environment. Scaling out, a path to modularization and decentralization, is done most often with blades.
Another trend is that enterprises no longer look to concentrate IT firepower in one fixed box. In recent years, significant increases in the power of commodity processors, the use of more decentralized applications (e.g., it doesn't make much sense to use a mainframe to support Web services), and virtualization on Intel platforms have all increased flexibility.
This plays into the hands of scaling out via blade approaches. Initially (i.e., way back in 2001) blades were used at the edge of networks for load balancing, DNS server applications, caching, and similar jobs, Dougherty says. The initial attraction was savings in space and overhead elements, such as cooling and powering. As blades got more powerful and sophisticated, they began to migrate from the edge toward the network center.
Today, it is not uncommon to use blades as the front end to a CRM, ERP, or similar application.
The scaling out vs. scaling up model can be expanded to a network-wide view by distributing server tasks into three tiers. While reality undoubtedly creates overlap, the three levels are fairly distinct:
- Tier one, served most often by rack-mounted servers, consists of applications that the end user "touches," says Paul Barker, vice president of marketing for RLX, one of the first vendors to bring blade servers to market.
- Tier two focuses on software integration, Web services, and similar applications. In this area blade servers predominate.
- Tier three servers typically are used for back-end tasks, such as database and transaction processing tasks. At this level, mainframes remain the undisputed king, Barker says.
The changes in what types of applications are assigned to what class of servers will continue to evolve as the underlying technology and applications change. This evolution will also be abetted by another trend: virtualization. Virtualization is the use of one physical machine to host multiple independent operating systems.
Server partitioning has long been a staple in the mainframe world. Now, companies such as SWsoft and VMware are bringing virtualization to the Windows and "Lintel" world. There has been much news from this growing industry segment during the past few months. "When you start virtualizing," says Dougherty, "it allows you to ... abstract things to a higher level."
This is, in essence, another tool for modularization, and as such it will accelerate the trend of blade adoption. Using blades and virtualization will enable decentralized systems to muster resources -- many of them "borrowed" via grids from computers that are momentarily idle -- to amass the kind of computing power formerly possible only from a mainframe. There is much interest in using blades running Linux for clusters, Dougherty says.
It has not been all smooth sailing for blades, however. The poor economy and the innate reluctance of CIOs and senior management in general to try new technology is slowing the adoption process.
There are also technical issues to work through. For example, the industry must master the overall management of widely diverse devices. If blades are being used collectively in place of a much larger single server, it stands to reason that there is a tremendous amount of scheduling, coordination and other complex management tasks that must be performed. The second element is that the speed at which data can be transferred between different machines can be a limiting factor in how fast they collectively operate. Ethernet, fibre channel, iSCSI and Infiniband are all approaches to trafficking data between processors.