IBM ups the ante with Power virtualization

Virtualization is one of the hottest buzzwords in IT today, but all the big news seems to revolve around x86-based systems and the companies that are virtualizing them, such as VMware, Citrix and Microsoft. But the original virtualization company – IBM – is somewhat quietly making strides of its own. IBM invented virtualization back in the 1960s to set up partitions within the mainframe, resulting in better hardware utilization. Big Blue has resisted the urge to develop an x86 hypervisor that would compete against VMware, but in 2001 the company began shipping a hypervisor for its Power processors, which offer considerably more horsepower than the x86 servers in your data center. At the recent VMworld in Cannes, VMware and its competitors advanced their vision of a more flexible IT model that joins private data centers to external “cloud” computing services, and lets workloads move seamlessly within data centers and from one data center to another. Specifically, VMware is promising to build a “Virtual Datacenter Operating System” a software layer that aggregates virtualized servers, storage and network resources into one big computing pool, and lets customers provision and manage all internal and external resources through the same interface. The VMware vision is dependent upon the use of x86 processors, which probably makes sense given their prevalence and price levels. But virtualization is relevant to numerous other types of systems. I wanted to see if IBM is working on anything similar, so I got on the phone with Satya Sharma, an IBM distinguished engineer based in Austin, Texas. It turns out Big Blue has a project called Ensembles, which involves a similar aggregation of virtual computing resources. “It’s about treating servers as a single resource,” Sharma explained to me. “Customers are telling us the cost of platform management and system management is escalating. Therefore, automation is a critical element. This idea is about automating operations across a pool of servers rather than managing them on a server-by-server basis.” Live migration of virtual machines from one server to another is a key capability, especially as data centers increase use of virtualization and put mission-critical applications on virtual servers. IBM began providing live migration when it released the Power6 chip in May 2007, and this technology will presumably play a role in Ensembles. Once available, the Ensembles technology will optimize the placement and movement of virtual machines - or logical partitions LPARs), a nomenclature borrowed from IBM’s mainframe virtualization. PowerVM, IBM’s hypervisor, ships with nearly two-thirds of Power-based servers and is more similar to mainframe virtualization than to VMware’s x86 variety, Sharma says. On Power, the operating system is aware of the hypervisor, allowing greater cooperation between the OS and hypervisor layers, he says. (With VMware, the OS is not aware of the hypervisor). Currently, the PowerVM hypervisor lets IT pros run up to 254 LPARs on a single 64-core server. With the upcoming Power7 processors, Sharma says IBM is “thinking of upping that number considerably” to take advantage of the extra capacity. PowerVM also lets IT pros add or remove processors, memory and I/O in a non-disruptive fashion, and allows the application of firmware patches without bringing down a system. “When you run a large number of LPARs on a system, there is never a good time to bring down a server for maintenance purposes. If I had 150 LPARs running on a system, it would be impossible to negotiate maintenance windows,” Sharma notes. Typically, IBM overhauls the PowerVM software each time it upgrades the Power chip, which happens every two or three years. Power7 is rumored for a 2010 release, but Sharma tells me IBM is planning some big improvements for the hypervisor even before Power7 hits the market. Memory virtualization will be among these improvements slated to be released before Power7. Right now, each virtual machine in a Power system has its own dedicated set of memory and cannot easily share it with another VM, Sharma says. IBM engineers are working on a system that will automatically share DRAM among numerous virtual machines, in much the same way the hypervisor already shares processing power among virtual machines. IBM is calling the technology Active Memory Sharing and has already released a beta version. Sharma is also promising “tremendous innovation” in the area of virtual I/O, which allows sharing of I/O adapters across partitions. IBM’s focus on memory and I/O virtualization was recently spotlighted in a Forbes article. “The key point is that you want to over commit memory, I/O and processing and run it at a high degree of utilization," Brad McCredie, chief architect of the Power6 chip, told Forbes. Sharma was reluctant to say when any of these technologies will come out, but it seems reasonable to expect that at least some of the new features this year. With virtualization being used on a large majority of Power servers, they will likely be highly anticipated.

Join the Network World communities on Facebook and LinkedIn to comment on topics that are top of mind.

Copyright © 2009 IDG Communications, Inc.