|
 |
|
General Information
|
Add To My Personal Library |
November 20, 2009
• Vol.31 Issue 28
Page(s) 39 in print issue
|
Servers Optimized For Virtualization
Carefully Configure Them For Best Performance
|
| Key Points
• The performance of server hardware has improved one to two orders of magnitude over the last decade, meaning that most systems are vastly underutilized when running a single OS image and application.
• Virtualization is the best way to maximize server utilization; however, while current quad-core CPUs can easily handle the processing load, other parts of the system can create bottlenecks.
• Servers for virtual infrastructure require plenty of RAM and I/O bandwidth. | | The inexorable and exponential improvements in PC and server hardware are nothing new. Yet early generations of x86 hardware barely kept up with software’s added demands, leading to something of a performance standoff: Systems became more powerful, but new application and OS functionality soaked up every CPU cycle. Despite software vendors bloating their products with often unused and unneeded “features,” those days of software-hardware parity are over.
According to Gary Chen, research manager at IDC, “a single OS installation has difficulty using all the power [of] a multicore system,” a situation that means the typical server runs at an average utilization of somewhere between 10 and 35%.
To combat this waste of resources, the industry has gone back to the future and taken a page from the mainframe era by parceling out servers into virtual slices and thus sharing the workload among many logically distinct machines. “Virtualization saved the server hardware market by giving it a way to use all that power,” says Chen. Yet running multiple VMs creates new system demands that can tax even the most modern hardware, so both chip and system vendors have incorporated new technology and features that optimize for virtualization.
 Key Attributes For Virtualization Platforms
According to Nik Simpson, senior analyst at Burton Group, there isn’t really a specific product category of “virtualization-optimized” server, as any system bought in the last two years already has the requisite CPU support for the most advanced features of today’s hypervisors. “Server virtualization is guiding the evolution of the commodity platform by transforming it into a platform optimized for one role: running multiple virtual machines on top of a hypervisor,” he says.
Paul Durzan, director of hardware platform marketing for unified computing at Cisco (www.cisco.com), says the three most important hardware attributes for virtual servers are adequate processor performance, copious amounts of system memory, and plenty of external I/O bandwidth.
As Simpson points out, Intel and AMD have long recognized the demands virtualization puts on a CPU and have optimized their server processors and motherboard chipsets for some of the unique workloads and memory management challenges presented when running multiple guest operating systems. For example, Chen notes that Intel’s latest Nehalem-based CPUs incorporate a number of technologies that improve memory use and I/O speeds, yielding a 150% performance improvement over previous processor generations, which translates into the ability to run 2.5 times the number of VMs per chip.
Given the I/O generated by many simultaneous VMs, optimized servers need plenty of internal bandwidth, exemplified by multiple 16-lane PCI-E slots, and a like abundance of external I/O capacity. Until recently, servers have met the latter requirement by using one or two quad-port Gigabit Ethernet LAN cards and separate Fibre Channel SAN adapters. With the maturing of 10GbE technology and concomitant price erosion for NICs and switches, Durzan suggests using Ethernet as a converged network for both data and storage networks—the latter via either the iSCSI or FCoE (Fibre Channel over Ethernet) protocols.
Harnessing and apportioning this bandwidth is another challenge for virtual servers and the focus of new hardware IOV (I/O virtualization) technologies. Although the PCI-SIG has two IOV standards under development that allow multiple VMs running on the same server to share I/O interfaces and PCI devices, Simpson advises proceeding with caution given their relative immaturity.
Form-Factor Choices
Buyers must also consider factors such as form factor and hardware redundancy. Simpson sees the market dividing into four segments: dual-socket systems in 1U or 2U, 4U quad-processor servers, and blade chassis.
Each has its advantages and disadvantages, with 1U and 2U systems trading off expandability and capacity for space and price, while 4U systems offer greater performance and VM consolidation ratios but take up more rack space. Although Simpson says “there doesn’t seem to be a single right answer,” he sees larger systems as an attractive option for those seeking to minimize the number of boxes they administer, while 2U servers are the market’s sweet spot.
Blades seem like a logical fit for a dense, consolidated VM server farm; however, Simpson notes that from a management standpoint, “they still look like a bunch of small, 1U servers.” He adds that “blade servers tend to be less scalable in terms of memory and I/O throughput, which translates to fewer virtual machines per physical host when compared to similarly priced rackmount servers. On the other hand, blade servers do offer more complete IOV schemes, reduced energy consumption, and simpler systems management.”
Yet Cisco’s Durzan feels blades offer some intriguing advantages, especially when coupled with the I/O virtualization technology. He says that with their shared power supplies and fans, blades are inherently more energy-efficient.
Deployment Guidelines & Recommendations
As server hardware has outstripped the ability for conventional OSes and applications to effectively utilize the available resources, Simpson’s recommendation for data center managers is to establish virtualization as the default server deployment model and don’t expect to reuse older hardware. He notes that current processors and hypervisors have eliminated the performance roadblocks to virtualizing most applications but cautions that a few issues remain, such as ensuring adequate memory and I/O capacity.
Chen agrees that virtualization has become the mainstream server application, and all newer systems have the processors and chipsets that support advanced VM features with minimal performance loss. He notes that as long as a system is supported by the virtualization vendor, it’s more important that buyers focus on system configuration, adding “at that point it’s a sizing exercise.” 
by Kurt Marko
Server Buying Recommendations For Virtualization
According to Burton Senior Analyst Nik Simpson’s research overview, titled “The Commodity Server Is Being Reshaped by Server Virtualization,” SMEs should keep the following server buying recommendations in mind when considering virtualization.
• Don’t use old hardware for server virtualization projects. Older hardware is likely to be missing key features in such areas as: processor extensions, processor performance, memory capacity and performance, I/O throughput, and the capacity to run multiple virtual machines.
• Identify the true roadblocks for server virtualization, which include support for some I/O interfaces and very high-performance requirements.
• Don’t get obsessed with performance loss for server virtualization. When considering the issue of performance loss, it’s important to remember that on the latest commodity server platforms, performance loss due to overhead associated with server virtualization is in the 5 to 10% range. And, many of the applications running on commodity servers are not performance-sensitive. (For example, they barely utilize the performance and scalability offered by the existing hardware.)
• Consider hardware-based IOV (I/O virtualization) sparingly. Hardware-based IOV looks attractive because it helps to remove some of the remaining performance loss due to server virtualization and provides useful features such as managed quality of service for each virtual machine, but IOV’s drawbacks include establishing a hardware dependency between a VM and specific hardware configuration and an immature software stack. |
|
|
