In the Field: Virtualization Capabilities on Unix and POWER5
In the Field: Virtualization Capabilities on Unix and POWER5
By VSM News Staff published: Thursday, October 07 2004
“This is a long time coming. And it’s game-changing.”
Keith Gauvin, an Enterprise SE with Expert Server Group of Bedford NH, talks with VSM about IBM’s recent announcement of virtualization capabilities on Unix and POWER5. Keith's extensive credentials include IBM P-Series Technical Specialist, P-Series AIX Advanced System Support Specialist, AIX System Administrator, P-Series Logical Partitioning Specialist, Red-Hat Certified, CompTIA Linux + Certified, Sun Certified Network Administrator, Sun Certified Enterprise SE.
VSM: Good morning Keith. IBM announced recently virtualization on Unix with POWER5. Can you tell us about that?
KG: Certainly. Virtualization on POWER5 is a continuation of the LPAR capabilities in POWER4. It allows us to use a more granular model, providing the ability to virtualize both the CPU and basic I/O subsystem. With POWER4 an entire CPU was required in a partition, and I/O devices had to physically exist in system.
The POWER5 chip is the next generation of POWER processors. POWER5 enables us to run multiple instances of an operating environment over a single chip. Each of these sub-CPU partitions is referred to as a micropartition. Currently up to ten instances of Linux or AIX can be supported by a single CPU.
Further enhancing this micropartitioning capability is the ability of the hypervisor to allocate unused CPU cycles from other partitions to those micropartitions that are under increased loads. You’re effectively moving into a model where not only are resources able to be virtual, they are dynamically allocated and assigned and can be moved based on chronology and workload intensity.
VSM: Tell us more about virtual I/O.
KG: In most environments the factor that constrains the design of a system the most is the I/O, the ability to move data in or out of the system. Typically I/O will span the spectrum from disk drives to network interfaces, anything that has to do with character, block or stream devices. A finite number of devices that can handle read or write capabilities is always going to be your limiting factor in the context of how many times you can either partition a full single system or the amount of data you can get in and out of a single system.
What IBM has done is create a virtual I/O engine that enables us to put another layer in the device driver stack, and allows us to provide many instances of a single resource used for I/O. For instance, we would use a network card, but we could then divide it up among several LPARs or micropartitions. Each of these instances would communicate through this single physical resource. Each instance is autonomous with respect to the assigned instance.
So by virtue of virtualization we can present this single physical entity more times than it actually exists to far more instances of an operating system that do actually exist.
VSM: Is the implementation of virtualization in the hardware and software?
KG: Yes and yes. You have to have hardware that’s able to handle the requirements of virtualization, which could be anything from increased bus speeds and PCI bridging to allowing CPU cycles to be shared and scheduled or re-queued – it’s more complex. Hardware design considerations are very much present.
From the software perspective, the applications themselves that live on an instance of the OS are no different, they are exactly the same as they would be on any other system. Conversely the operating system instance itself does have to have some features built into it, primarily components that allow it to live in a dynamically virtual environment. To have its resources moved and allocated dynamically can be very complex at the software level.
Currently there are three supported OS versions on the IBM POWER5 lineup. There are two distributions of Linux, SLES9 and Red Hat Advanced Server 3, as well as AIX 5.3. AIX 5.3 supports SMT, which is Symmetric Multi-Threading, as does SLES9. I feel it is important for listeners to know that IBM has over 600 Linux programmers and contributors on staff currently. They are working very closely with the open source community to make sure that POWER5 is tuned for Linux and that Linux in turn is tuned for POWER5.
VSM: Is this available right now?
KG: Absolutely. This is prime time, front and center. This has been around in the mainframe zSeries for 10+ years. With bandwidth limitations we saw computing move out towards the end users with powerful workstations. Now with gigabit connections and high-speed bus fabrics we will see a continued return to the centralized computing model. It has been refined and re-platformed for HPC RISC chips and Unix.
This is a great technology to group together all of those disparate older systems into a single HPC footprint. Case in point: Perhaps you are running four dedicated WebSphere systems, with 275MHz CPUs and 4GB of memory. Using rPerf and a sizing analysis we can actually take those servers that live out in the physical environment and get a one-for-one or better performance return in a virtual environment.
We can consolidate them into a POWER5 platform and quite literally get the same or better performance than that box could be providing if it still existed out in the physical environment. You’re not giving anything up by reducing your footprint at this point in the game. They’ve gone to great lengths in making sure that the performance is adequate, the performance is dynamic, and it’s there to go ahead and do this consolidation at that level.
VSM: Is the implementation of UNIX virtualization straightforward, or is it tricky, with lots of subtleties and a need for specialists?
KG: Professional services are not something I would want to shy away from with respect to projects or implementations of any magnitude or importance. Consolidations require expertise and skills in assessing, planning, project management and execution from multiple departments and/or consultants. Add to that a new and involved strategic virtualization initiative and you are going to want some direction, training and accountability.
Unless you have extensive experience with POWER4 and some experience with POWER5, or extensive experience with POWER5, virtualization in a Unix environment is not an off-the-shelf solution. There’s advanced understanding required, of resource allocation, of existing technology as it pertains to IBM and as it pertains to partitioning, of the VI/O engine that will live on an actual instance of an OS.
So it is not straightforward out of the gate, it is something you would most certainly want to engage the help of experienced professionals to guide you through the process and architect it.
Ironically some of the more logistical components, which may not be obvious up front, are really where most of these consolidations may become difficult. So proper planning and appropriate technical resources are all very important.
VSM: You’ve mentioned server consolidation. Are there any other business drivers that would lead a customer to implement virtualization?
KG: The application for POWER5 fits nicely into the consolidation paradigm, however that is not the only time to consider POWER5. POWER5 is a very fast RISC-based chip. It has advanced features like Symmetric Multi-Threading. It’s very robust in that it scales, it will maximize the available performance for Linux and AIX.
So any application where people are looking to deploy applications that are very responsive, very quick, with high I/Ops and very high SLAs, POWER5 is most definitely one of the best avenues to look at.
VSM: Will virtualization make Linux more attractive as an add-on operating system?
KG: It will allow developers and engineers to work in an environment that comes at a lower cost to the organization as a whole, where they could be given partitions to develop and create applications as required without needing the organization to procure new hardware. So, yes.
VSM: Can you talk about the communications between partitions?
KG: In any network environment a system must maintain some form of network connectivity and typically that’s done through Ethernet at this point in time, that tends to be the predominant technology.
Ethernet as a technology has limitations in bandwidth and configuration requirements, and what often happens in a virtualized environment, as on the Intel platform, is that you no longer have to travel over a physical wire or through a physical host bus adapter to communicate this protocol. You actually are going directly over the bus to the next partition.
So oftentimes to meet the requirements of operating systems and their ability to handle certain data rates we will simulate that through the bus. We would then throttle that back to a 10/100 or 10/1000.
In other words in virtual partitions, or partitions in general, or micro-partitions, these are actually isolated instances, and you are going to have to talk via network media. It’s not as if they just share all common file systems. You’re still going to have to set up IP interfaces, IP protocol stacks, device drivers and instances.
You’ll have some that are shared through the VI/O, and you can have some internal to other partitions only on a separate network. So it would be a segregated network in that only those partitions would talk to each other on that network.
VSM: Is it a fact that in addition to consolidation that because of virtual I/O, companies save money? Instead of needing a host bus adapter or more likely two for each partition, because they are able to virtualize it, they don’t have the expense of an HBA per partition anymore?
KG: That’s correct. In most instances software is purchased on a per-processor license agreement. That has the potential save you money. In addition to that you’ve got a reduced number of physical I/O devices that are required.
In a traditional environment, as you grow and scale you need to add more devices and or more instances of the OS. POWER5 helps to eliminate that. You do not need to do that anymore. You’re just going to create another virtual instance of a device that is being underutilized.
VSM: If I had ten partitions, how many HBAs would I have to purchase?
KG: Depending on the expected load, one or two.
VSM: That’s quite a savings.
KG: 10:1.
VSM: So in addition to the many savings from consolidation, there’s a savings right in the machine from a hardware standpoint.
KG: Definitely. There’s less power needed to drive these machines because there are fewer components required. There’s less involvement from a computer engineer, an onsite repair service, because there are fewer components, a smaller footprint.
VSM: Virtualization on Intel has been all the rage the past couple of years. Because of the stability of Unix, is it a more attractive platform for virtualization? You’re putting all your eggs in one basket, but because it’s a Unix basket is it inherently a safer and more stable way to implement virtualization?
KG: I would say yes, I agree with that. It’s a stable platform. It allows virtualization to occur in a stable fashion, arguably more stable than in a Windows environment. Hence the use of Linux as a virtualization kernel on the Intel side for enterprise based applications.
VSM: For the Linux and the Unix world, is this a big deal?
KG: This is a long time coming. And it is game-changing. It’s now putting the number of services per administrator as the metric to measure your IT infrastructure, as opposed to the number of servers per administrator. That’s revolutionary, because originally we would measure an IT administrator’s relative skill set based on the number of physical boxes he or she could maintain.
And those administrators with different platform expertise had different expected numbers. A Sun admin may be able to do 30 or 40, where a Microsoft admin would be expected to do 10 to 20. That’s based on the number of problems, support required, updates, packages, problems of that nature.
Essentially now we’re taking those physical boxes out of the equation. We no longer need to know how many physical boxes this person manages, we need to know how many on-demand services our administrators are going to be able to handle. And that means a single box with multiple instances of an operating system.
That’s a new way of looking at it. That falls into the services-on-demand model that I think we’re all striving for. Where you’re going to plug into an outlet and get those services you want. It’s a more efficient model, without question, the natural evolution of IT infrastructure.
VSM: If you were looking at this from the perspective of a CTO or a CIO, what other implications are there for the enterprise?
KG: If I were to put myself in the shoes of a CIO or CTO, my main concerns would be cost, return on investment, rate of depreciation, continued support. A reliable roadmap, expected support, changes, upgrades that are on a nice drawn-out interval. And those are all things that have been kept in mind as IBM has continued to develop its POWER4 and POWER5 E series servers. Those are all questions that are answered with a check in the Yes box when we’re looking at POWER5.
VSM: From your perspective, that would make it less risky?
KG: Absolutely. Less risky, lower cost, predictability, scalability. Hands down, goes to IBM.
