The Benefits of POWER7+ and PowerVM over Intel and an x86 Hypervisor Howard Anglin rhbear@us.ibm.com IBM Competitive Project Office May 2013
Abstract...3 Virtualization and Why It Is Important...3 Resiliency for Mission Critical Workloads...4 Power Systems PowerVM and Virtualization Inefficiencies...5 Workload Density...6 Capacity on Demand (CoD)...8 Dynamic LPARs (DLPAR)...10 Active Memory Expansion...10 Migrating Virtual Machines between Hosts...11 Live Partition Mobility (LPM)...11 Virtual Machine Mobility...12 Summary...13 Page 2 of 14
Abstract The popularization and advancement of virtualization technologies have resulted in the consolidation of many server workloads onto more powerful machines. Increasingly heavy demands are placed on IT environments to support the resulting virtualized workloads where dynamic allocation of computing resources, namely CPU, memory and I/O, are now required. IBM PowerVM can effortlessly accommodate many different types of virtual machine (VM) workloads without the need to constantly relocate demanding VMs to satisfy resource requirements and Service Level Agreements (SLAs). This article will show that PowerVM technology delivers more advantages than an x86 based virtualization solution for virtualized workloads and cloud computing. It will also highlight some of the benefits PowerVM brings to organizations planning to implement virtualization technologies in their data centers. Furthermore, the reader will see that Intel-based servers and x86 hypervisors lack many key features needed to efficiently manage and effectively support virtualized environments that are essential to a truly dynamic IT infrastructure. PowerVM and POWER7+ are key enablers for cloud computing with built-in features including capacity on demand (CoD), active memory management, shared processor pools and the ability to accommodate high workload densities consisting of a blend of high and low priority workloads on a single server. Virtualization and Why It Is Important Virtualization is a method by which hardware resources are abstracted out to multiple logical hardware resources allowing multiple virtual resources to be hosted on the same physical hardware. The layer of software that serves up virtual instances of the physical hardware resources is known as the hypervisor. The virtual instance of physical hardware, the virtual machine (VM) has all the characteristics of a functional physical server and provides processing, memory, storage and networking. The general representation of virtualization in Figure 1 illustrates a physical server (dotted line) hosting several guest virtual machines that has access to memory, CPU, network and storage resources via the hypervisor. Page 3 of 14
Virtual Images Hypervisor CPU Memory LAN Storage Figure 1: Virtualization - hypervisor and hosted virtual machines Virtualization is a key enabling technology for data center consolidation and cloud computing environments where physical resources are shared and made available through an appropriate service delivery model. Consolidation workloads on fewer more efficient systems can dramatically reduce capital operating expenses in IT environments. Resiliency for Mission Critical Workloads Increasing workload density per physical server may increase the risk of disruption or may violate the terms of a Service Level Agreement (SLA). Power Systems technology leverages IBM s leadership in performance, virtualization and security to sustain non-disruptive growth for mission critical highly available workloads. PowerVM is embedded into the hardware at the firmware level, unlike other hypervisors, helping to ensure one of the industry s most secure platforms. Hardware cost savings may not be the only motivation behind the virtualization of workloads. Certain types of workloads, due to the nature of the workload itself, may not see dramatic cost reduction. Complex workloads and critical applications such as analytics, ERP or high performance transactions - are instead being made more dynamic by making them easier to transfer between hosts, creating faster time to value and ease of duplication through virtualization. PowerVM really shines in this area with built-in capabilities to dynamically reallocate resources to critical business workloads should these workloads need more capacity. The entire Power virtualization ecosystem is owned by IBM, from the hardware to the operating system and IBM has optimized the stack to provide efficiency and resiliency advantages to the hosted workloads. A good-enough solution hosted on x86 servers will Page 4 of 14
quickly run into resource problems as the demands for quality of service increases. Large workloads with ever increasing demands and efficient consolidation require large capacity servers that x86 based solutions cannot provide. Even at the high end, the enterprise class virtualization solution from a competitor and its hardware provider may not satisfy businesses growth and operational needs of the customer. The aforementioned enterprise solution supports a maximum of 64 virtual CPU (vcpu) and 1TB of virtual RAM (vram) per VM while PowerVM partitions are scalable to 256 vcpu and can use all available memory on a server, up to 16TB. Power Systems PowerVM and Virtualization Inefficiencies PowerVM is the logical choice for a highly scalable virtual platform that can virtualize more workloads than the competition. PowerVM supports hardware with up to 16 TB of memory (all 16 TB can be used by LPARs) while an x86-based hypervisor can top out on hardware support with as much as 2 TB of memory, where a maximum of 1 TB can be used by virtual machines. All the relevant metrics required to virtualize critical or complex workloads are offered by IBM to infrastructure planners. PowerVM supports up to 1000 partitions per server (Power 770 795 ) compared to the alternative that only supports 512 VM per x86 host PowerVM supports up to 1024 threads per partition (LPAR) vs. 64 threads per VM hosted on an x86 server Some rivals can automatically hot-add CPU and memory resources to virtual machines; the only problem is that resources cannot be returned to the resource pool without rebooting the VM. Power Systems effectively address virtualization operational inefficiencies by reducing the need for migrations to resolve resource contention, itself a resource intensive operation, and still sustain higher VM densities. The following section highlights some of the features of Power Systems and emphasizes key advantages over competitors: Workload Density Capacity on Demand (COD) Dynamic LPARs (DLPAR) Active Memory Sharing Page 5 of 14
Workload Density Workload density refers to the number virtual machine instances that can be effectively supported per physical server. Workload density is determined by various important factors; the size of virtual machine workloads versus the capacity of the host, the ability of the hypervisor to accommodate the workloads without disruption and the variability of the workload demand per VM. In our tests PowerVM on POWER (Power 730) shows a 2.2x workload density advantage over an x86 hypervisor on Intel Sandy Bridge for a medium web facing workload and up to 2.5X advantage for heavy workloads. Figure 2: Virtualization density - VM size per server for a given (workload) user interaction per second PowerVM on POWER7+ servers offer better performance and workload consolidation ratios for virtualized environments than a hypervisor on Intel. The newest Power servers can sustain very high host server utilization while simultaneously running low priority and high priority workloads. Unlike the solution on Intel servers, Power can accommodate mixed priority Page 6 of 14
workloads and maintain higher server utilization without the need to constantly move VMs around. Figure 3: Virtualization on Power enables mixing of high and low priority workloads without penalty 1 To satisfy the requirement to host varying workloads on a physical server, hosting platforms must be able to differentiate between workload types and properly allocate shared resources. In the interest of maximizing the utilization of the hosting platform, low priority workloads must be able to share resources with high priority without adversely affecting the other. Other hypervisors permit leakage, where low priority workloads take away resources from high priority workloads, which minimize how dense workloads can be stacked. For the cost conscious customer, the cost per workload is a key metric for virtualized platforms; the higher the consolidation ratios the lower the cost per workload. A single server with a large pool of shared processors can run more virtualized workloads than several smaller servers with the same total number of processors. 1 Intel Server Specs IBM System x3850/3950 X5 (7143); 40-core/4 socket Intel Xeon E7-4870 (Westmere) 2.4 GHz, Hyperthreading On Power 760 Specs IBM - Power 760+ Express 3.4 GHz: UNIX: (4ch/48co Dual Chip Module); - Two 6-core chips in each socket = 12 cores per socket; Minimum entitlement given (0.05/core) to the donor LPARs, the balance to the priority LPARs. Page 7 of 14
Figure 4: Larger Servers with more resources make more effective Virtualization Platforms [workload - 3000 user interactions per second of a banking application] Live Partition Mobility (LPM), which will be looked at in detail later in this document, is a Power Systems based solution that moves running LPARs form one physical server to another without disruption or noticeable degradation of service. Users can rely on LPM as a migration tool to consolidate running LPAR workloads on POWER7+ based servers for high density host server operation. Consequently, bigger servers with capacity to run more workloads can be driven to higher average utilization levels without violating service level agreements, thereby reducing the cost per workload. Workload consolidation and reduced cost per workload are desirable goals for any IT department, however high density hosting is heavily influenced by the capacity of the host servers. Power-based servers can supply the type of high capacity servers required by large workloads with the highest number of available cores at 256 cores on Power 795 while x86 based servers top out at 32 cores. Capacity on Demand (CoD) Certain workloads grow in resource requirements where scaling up or dynamically redistributing resources to that VM can no longer address. The workload will no longer fit on Page 8 of 14
the small VM and may need not only bigger VM but a bigger physical server. This scenario appears to be the perfect use case for Live Partition Mobility or other VM mobility solutions. Nevertheless on Power, an alternative is to simply increase capacity on the underlying hardware (processors and memory) without disturbing the hosted workload with Capacity on Demand (CoD) 2. Figure 5 : Peaks can often be difficult to predict, especially their magnitude The Power architecture can dynamically add capacity to physical systems as needed, for example by either of the two following means: 1. Activate dormant processors or memory units within your system for a number of days, as business peaks demand Elastic CoD 2. Provide additional processor capacity on a temporary basis within the shared processor pool in increments of processor minutes Utility CoD Dynamically allocating and de-allocating resource to particular VMs is a far better option than moving the VM itself. Moving your critical workload increases the opportunity for error conditions to occur, from the network during transport to unknown conditions at the target host. On demand compute resource reallocation for VMs is more efficient and can lower overall costs through significant software license savings as additional licensing costs are limited to only extra resources used during peaks. Consider this analogy of an office building with an ongoing electric utility need where the current residents are approaching the supplier s 2 www.ibm.com/systems/power/hardware/cod Page 9 of 14
threshold. One solution is to relocate a portion of the residents across town in another building. A better and less labor intensive solution is to first choose a facility with the ability to quickly respond to varying and increasing electrical resource needs. Budget conscious IT managers and businesses need the flexibility to quickly respond to new demands without over provisioning resources - paying for resources before actually using it - or to choose a less efficient and potentially risky path of unnecessarily moving the workload. Furthermore, migration of complex business-critical applications is not always possible during a resource shortage. Dynamic LPARs (DLPAR) A primary reason for VM mobility is the need to secure or guarantee resource requirements of a VM and maintain SLAs. This means identifying the guest VM and host server with the need, and then identifying a target server, moving the VM to the target while preserving state and inflight memory information. PowerVM on POWER7+ can easily guarantee resources with DLPAR (LPARs the VM equivalent in the x86 environment) a facility available since POWER4. DLPAR allows the system to dynamically move shared memory and processors in the resource pool between LPARs, increasing and decreasing resources on the fly without rebooting VMs or the need to interrupt operations. Rather than rely on a resource scheduler feature for automatic VM load balancing - placement and subsequent migration of VMs between hosts to attain the best use of resources - simply dynamically reallocate resources to the appropriate LPARs instead with PowerVM on Power Systems. Relying on virtual machine migration as a way of managing VM resource needs can be viewed as a solution to address the technical limitations of a non-powervm virtualization solution. Instead of responding instantaneously to the needs of the hosted VMs, the x86 hypervisor tool relocates the offending VM to a host that may be able to alleviate the problem - measured in seconds or minutes versus instantaneously with dynamic LPARs. The time to complete a migration operation is typically longer than allowed by SLAs; clearly this approach is not the most effective way to manage resources in a data center. With PowerVM and Power Systems high performance CPU cores, resources can be scaled up and back to handle pretty much any VM workload that comes its way, scaling from a fraction of a CPU up to 64 CPUs. Aside from planned down time, the need to actually move LPARs around is significantly reduced with POWER. Active Memory Expansion Page 10 of 14
Active Memory Expansion (AME) reduces physical memory requirement for an LPAR and enables memory expansion by effectively inflating real memory size by up to 125% on POWER7+ without additional physical memory. The latest POWER7 processor leverages onchip hardware acceleration to expand memory beyond physical limits. AME boosts workload density allowing more workloads per partition, effectively run more workloads per server with less processor overhead. Figure 6: Active Memory Expansion enables memory expansion up to 125% on POWER7+ systems Migrating Virtual Machines between Hosts Live Partition Mobility (LPM) The PowerVM and Power System features that increase virtualization efficiencies may not completely eliminate the need to migrate workloads between servers. Planned outages and the need to refresh server hardware to say upgrade to the latest POWER7+ servers can be expected. Live Partition Mobility on IBM POWER6, POWER7 and POWER7+ servers allow users to migrate Logical Partitions running AIX, IBM i and Linux operating systems (since POWER4 for AIX only systems) and their hosted applications from one physical server to another. The migration transfers the entire virtual system environment, including processor state, memory, attached Page 11 of 14
virtual devices, and connected users. The process allows the movement of a running LPAR from one physical machine to another without disrupting the operation of the applications running in that LPAR or the infrastructure services. Live Partition Mobility contributes to the goal of continuous availability, as follows: Reduces planned down time by dynamically moving applications from one server to another Responds to changing workloads and business requirements when workloads are moved from heavily loaded servers to servers that have spare capacity Reduces energy consumption by allowing IT to easily consolidate workloads and power off unused servers In addition to LPM, another method available from IBM for VM mobility is Live Application Mobility (LAM) used to move Workload Partition or WPARs. WPARS are virtualized workloads that are hosted by an LPAR, each WPAR obtains a regulated share of system resources (AIX kernel, libraries, CPU and I/O, etc.) but all share a single instance of AIX. Workloads that share the same OS version and patch level can be consolidated as separate WPARS with a partition and used resources more efficiently than LPARs. Virtual Machine Mobility Figure 7: Migrating a WPAR to a different AIX instance with LAM There are now options from other virtualization vendors where it is possible to move virtual machines from one host another without disrupting the operation of the virtual machine. Rivals may offer a similar solution to Live Partition Mobility for live VM workloads. Unlike PowerVM, where any POWER6 or later based processor supports LPM, it is important to select hardware with CPU Page 12 of 14
compatibility for an x86 hypervisor and resource management vendor, to ensure a successful migration. PowerVM LPM supports live migration over many generations of Power hardware. Figure 8: VM migration - Moving an x86-based VM to a new host Summary As virtualization matures, clients now seek the right technology provider than can efficiently and simply deliver on the promises of virtualization and cloud capabilities. With the industries most reliable and advanced systems - POWER7+ servers, leadership virtualization efficiency and management, PowerVM leads x86-based add-on virtualization products for both large Enterprise applications and mixed workload environments. The primary approaches that are used by an x86 vendor to maintain an environment of virtualized workloads that are no longer running smoothly are often inefficient or otherwise inflexible. The VMs in a migration consume large networking resources and is a relatively time consuming to process for multiple VMs, even for enterprise level resource management solutions for x86 hypervisors. Additionally redistributing resources between VMs is inflexible as VMs need to be restarted in order to release resources back to the resource pool. PowerVM on the other hand has the built in flexibility to respond to these needs with several very efficient approaches and alternatively with LPM when necessary. PowerVM leads in virtualization technology against x86 hypervisors in scaling (support of large partitions and platforms), dynamic reconfiguration, security and support for multiple operating systems on a single sever and the many built in resiliency features make it an ideal solution for a virtualization platform. Page 13 of 14
Copyright IBM Corporation 2013 IBM Corporation Software Group Route 100 Somers, NY10589 USA Produced in the United States March 2013 All Rights Reserved IBM, the IBM logo, PowerVM, POWER7, POWER7+ and Power Systems are trademarks or registered trademarks of International Business Machines Corporation in the United States, other countries, or both. Intel, Westmere and Sandybridge are trademarks or registered trademarks of Intel Corporation. Other company, product or service names may be trademarks or service marks of others. The information contained in this documentation is provided for informational purposes only. While efforts were made to verify the completeness and accuracy of the information contained in this documentation, it is provided as is without warranty of any kind, express or implied. In addition, this information is based on IBM s current product plans and strategy, which are subject to change by IBM without notice. IBM shall not be responsible for any damages arising out of the use of, or otherwise related to, this documentation or any other documentation. Nothing contained in this documentation is intended to, nor shall have the effect of, creating any warranties or representations from IBM (or its suppliers or licensors), or altering the terms and conditions of the applicable license agreement governing the use of IBM software. References in these materials to IBM products, programs, or services do not imply that they will be available in all countries in which IBM operates. Product release dates and/or capabilities referenced in these materials may change at any time at IBM s sole discretion based on market opportunities or other factors, and are not intended to be a commitment to future product or feature availability in any way. Page 14 of 14