|
 |
|
General Information
|
Add To My Personal Library |
December 26, 2008
• Vol.30 Issue 52
Page(s) 1 in print issue
|
Beyond Blades
New Hardware & Software Technologies Let Enterprises Build Cloud-Like Server Farms
|
The landscape for industry-standard servers has been relatively bland for the last few years, with new products focusing on incremental improvements in performance or density at the same or lower price. Blade servers have led the way in high-density systems and have become a favored platform for virtual server deployments.
Yet the combination of blades and virtualization is rapidly evolving by incorporating new hardware and software to enable so-called cloud-like infrastructure. Research group Gartner has identified 10 strategic technologies and trends to watch in 2009. Among these is the maturation of traditional blades and virtualization that will enable more finely grained provisioning of various server resources, including everything from processors and memory to storage and I/O. Gartner sees a world evolving to “simplify the provisioning of capacity to meet growing needs. The organization tracks the various resource types—for example, memory—separately and replenishes only the type that is in short supply.” The benefit is “higher utilization because of lessened ‘waste’ of resources that are in the wrong configuration or that come along with the needed processors and memory in a fixed bundle.”
There are a number of hardware and software trends building the foundation for such utility-like provisioning. While blades and other high-density server configurations are a key element, server virtualization software and new I/O virtualization appliances will also play a major role in delivering on the vision of cloud-like infrastructure.
 Server Density Increases
Blade servers have evolved from crude repackaging of standard server motherboards into custom-engineered systems designed to house a variety of processor, storage, and network modules in a compact, easily deployed chassis. State-of-the-art blade chassis host up to 16 half-height blade servers in a 10U system, with each server sporting two quad-core processors. The ability to pack 128 processing cores in a 10U form factor is impressive, yet blades are not the only means to a high-density server farm.
1U server chassis have been a mainstay in most data centers, yet their wide and thin design means they aren’t typically as dense as their blade counterparts. Recently, several vendors have introduced some creative packaging to the standard 1U chassis that has leapfrogged blades as the density champion. These systems typically incorporate two daughterboard modules, each supporting two quad-core CPUs, ample RAM sockets, and four drive bays in a standard 1U chassis.
Virtualization Software Enables Fine-Grained Provisioning
Today’s servers pack so much horsepower in such a small space that a pressing problem has become how to efficiently deploy such an abundance of resources without over-provisioning or otherwise wasting capacity. Server virtualization has become the de facto solution for improving CPU utilization and consolidating multiple applications onto a high-density blade chassis or server cluster. While virtualization software has gotten better at allocating processor resources, Taneja Group Senior Analyst Jeff Byrne says that virtualizing memory across multiple VMs remains an unsolved challenge. Likewise, he notes that creating a cloud-like infrastructure also requires virtualizing I/O and storage.
Memory management in a virtualized environment is still relatively crude, with most hypervisors supporting only a fixed allocation of RAM to each VM. Perhaps signaling a trend in memory allocation technology, VMware touts its overcommitment or “memory ballooning” technology that allows the VMM (virtual machine manager) to allocate more memory to running VMs than actually exists on the physical server. Of course, not every VM can use its full allocation simultaneously; however, because the VMM can dynamically change memory allocation as needed, and because most applications will rarely need their full allotment, the scheme makes more efficient use of existing RAM. Byrne expects other variants and improvements in VM memory management from other vendors.
I/O virtualization is another technology key to a more efficiently provisioned, cloud-like infrastructure. The problem, according to IDC analyst John Humphreys, is that “Current server I/O architecture is extremely inflexible because device addresses are typically associated with the physical device. This significantly drives up cost and complexity in the data center, especially as the number of servers grows and as each device requires a host of LAN and SAN connections that are tied to a specific machine.”
Arun Taneja, founder of the Taneja Group, points out a number of technical approaches to I/O virtualization, ranging from the creation of high-speed networking fabrics (typically using Infiniband or 10GbE) to switches using external PCI Express buses.
A counter-trend to the increasingly granular resource provisioning enabled by virtualization is so-called server aggregation. Taneja believes that the ability to gang servers together—in essence creating a large virtual machine out of many physical systems—while still a niche technology should get more important as enterprises base their infrastructure on small, standardized, dual-processor systems. Aggregation lets applications “see” multiple blades or 1U servers, say 64 processors or more, as a single machine—conceptually similar to what Intel’s Hyper-Threading does for multicore CPUs. Gartner VP and Distinguished Analyst Tom Bittman is somewhat dismissive of the need for aggregation, noting that Katana pioneered this concept several years ago and couldn’t find a market. The company subsequently abandoned the technology and rebranded itself to focus on conventional virtualization products.
The Future Of Enterprise Clouds
Building an enterprise cloud computing environment is becoming feasible thanks to increasingly capable server hardware, shared storage networks, virtual I/O technology, more sophisticated virtualization management software, and a services-oriented application architecture. Blade servers will undoubtedly play a major role in building the dense server farms required but face competition for other server packaging architectures. VMs will also need to be augmented by special-purpose I/O fabrics and SANs; all glued together via what Bittman terms a meta operating system managing a distributed fabric of virtualized resources. 
by Kurt Marko
|  This figure illustrates a typical architecture for a shared, virtualized I/O fabric for a blade server.
Source: “Windows Server Scalability And Virtualized I/O Fabric For Blade Server”; Presentation to WinHEC 2006; Chris Petty, CTO, NextIO; Son VoBa, program manager, Microsoft; May 2006. |
|
|
