Abstract
AIX offers two methods of vertical server consolidation: workload management with Workload Manager; and partitioning, of which the most recent development is shared processor logical partitions (Micro-Partitioning technology) with Partition Load Manager. This technote compares these two approaches.
Contents
Introduction
As we discussed in the technote Server consolidation on IBM pSeries Systems (TIPS0306), there are two approaches to server consolidation: vertical consolidation, where many servers are consolidated on to fewer, more powerful servers; and horizontal consolidation, where multiple servers share a common workload, providing enhanced scalability and reliability. Within vertical consolidation, AIX provides two options: Workload Manager (WLM) and Logical Partitioning (LPAR).
With the introduction of shared processor logical partitions (partitions using Micro-Paritioning technology) and Partition Load Manager (PLM), partitions are approaching the flexibility and granularity of WLM classes in their responses to changing load, while providing the additional security of separate operating systems. The sections below compare WLM classes and partitions using Micro-Paritioning technology in terms of their ability to dynamically provision resources (CPU, memory and I/O) to applications, and the features they provide. Partitions using Micro-Partitioning technology are not necessarily smaller than 1 CPU, but they can be given CPU entitlement in fractions of 0.01 CPUs (1.75 CPUs, for example). We assume that WLM is configured on dedicated processors.
For further documentation on the products and server consolidation, see:
- IBM Redbook: AIX 5L Workload Manager (WLM), SG24-5977
- IBM Redbook: Introduction to Advanced POWER Virtualization on IBM p5 Servers, Introduction and basic configuration, SG24-7940
- IBM Redbook: IBM e-server p5 Virtualization Performance Considerations, SG24-5768
- IBM Redbook: Server Consolidation on IBM pSeries Systems, SG24-6966
- Redpiece: LPAR Heterogeneous Workloads on the IBM pSeries 690 System, REDP-0425
Requirements and configuration
| WLM | Micro-Partitioning + PLM |
| WLM is provided free with AIX. | Micro-Partitioning and PLM are provided as part of the advanced POWER virtualization feature for AIX, which is a chargeable option. |
| WLM is installed by default. No additional hardware is required. | The managed server must have an HMC. LPARs must be defined and installed, and have Resource Management and Control (RMC) connections to the PLM server. The PLM server must be separately installed. |
| WLM classes, tiers, limits, shares and rules must be manually configured. | POWER Hypervisor (PHYP) entitlements, and PLM shares and capping must be manually configured. |
Allocation and separation
| WLM | Micro-Partitioning + PLM |
| All processes within an operating system (OS) are assigned to a class. | All processes run within a partition. |
| All classes run within the same OS. An OS crash will stop all the classes. | Partitions run separate OSs. An OS crash in one partition will have no effect on the others. |
| A process in one class can start a process in another class. | A process in one partition can only start a process in another partition using network communication. |
| Resource sets can be used to restrict a class to particular CPUs. | The administrator has no control over which CPUs in the shared pool are used by a particular partition. However, LPARs can be grouped so they only compete against others in the group. |
Performance overhead
| WLM | Micro-Partitioning + PLM |
| WLM is built into the definition of a process. Once running, the overhead is minimal. | Resource Management and Control (RMC) services gather and export the system status. The RMC daemon also processes reconfiguration (dynamic LPAR) requests from the HMC. |
| WLM can significantly increase the boot time of an OS if there are a large number of disks attached. | The RMC services are always started on boot. |
| Only one OS is required. | Each partition must have its own OS. |
| Dedicated partitions are the 'default state' against which partitions performance is measured. AIX 5.3 on POWER5 has set a number of benchmark records. | The performance penalty of sharing processors depends on factors such as the size of the partition and the number of other partitions running. The causes of the overhead are due to increased context switches and loss of cache. Virtual processors can be dispatched on different physical CPUs, or the cache may be overwritten by another partitions, losing the benefits of affinity. |
Resource entitlement
| WLM | Micro-Partitions + PLM |
| Classes can have maximum, minimum, and target resource entitlements. A class may be given less than its target, if all classes are under heavy load. It will only be given less than its minimum if it cannot use the resources, or if a higher tier class (see "prioritization") takes all the resources. | Partitions can have maximum, minimum, and guaranteed resource entitlements in the PHYP. A partition will only be given less than its guaranteed amount if it cannot use the resources assigned to it. It will never be given less than its minimum entitlement. |
| Target entitlements are known as shares. The resources given to a class are determined by its share divided by the total number of shares for active classes. An active class is one with running processes. | Partitions are assigned a share in PLM. The resources given to an LPAR are determined by its share divided by the total number of shares for active LPARs. PLM will override the PHYP's normal distribution of these additional resources. |
| A class with a maximum entitlement of 100% can use any free resources on the system. | An uncapped partition can use any free resources on the system, as PLM will increase a partition's virtual processors in order to exploit additional CPUs. |
| I/O throughput can be controlled. I/O resources can be shared between classes. | I/O throughput is not controlled. I/O resources can only be shared through a VIO server. PLM cannot move I/O resources between partitions. |
| The sum of the defined minimum resource entitlements of all the classes cannot exceed the total capacity of the system, even if some classes are not active (have no processes running). | The sum of the defined minimum capacity entitlements can exceed the total capacity of the system as long as not all the partitions are started. |
Prioritization
| WLM | Micro-Partitions + PLM |
| Classes can be put into tiers. Processes in a lower tier class will only run if no higher tier processes are running. Higher tier classes, therefore, cannot be limited by lower tier classes, but lower tier classes can be starved. | PLM has no concept of the importance of a workload beyond the share setting (see "resource entitlement"). Running a lower priority partition will limit the resources available to a higher priority partition because the lower priority partition will still use its guaranteed entitlement. However, lower priority partition cannot be starved. |
| Processes can be started, and classes activated, even if they cannot achieve their minimum entitlement. | New partitions will not start if their minimum requirements cannot be met. |
Speed of response to changing load
| WLM | Micro-Partitions + PLM |
| There is no latency associated with a class using additional CPU. | There is a latency associated with dynamically adding virtual processors. Furthermore, if a high number of virtual processors are made permanently available instead, a performance overhead is incurred. Additional entitlement (up to 100% of a partition's virtual processors) can be added without delay. |
| Monitoring is constant. Access to a class's resources is provided on a per-minute basis (as long as the class can use its full entitlement). | Monitoring is based on 10 second intervals. By default, a threshold must be reached 6 times in order to trigger a dynamic LPAR event. Entitlement changes are made only when an event is triggered, but excess capacity is distributed constantly (based on shares). |
Conclusion
WLM still provides a greater degree of control and granularity, and classes are still more dynamic in their response to changes in load than a partition using Micro-Paritioning technology, although these differences are becoming less noticeable. By running separate operating systems, Partitions provide an additional degree of separation with clear advantages for availability. PLM can also run with dedicated partitions, avoiding any performance overhead of paritions using Micro-Partitioning technology, but reducing the granularity of control still further.
Special Notices
This material has not been submitted to any formal IBM test and is published AS IS. It has not been the subject of rigorous review. IBM assumes no responsibility for its accuracy or completeness. The use of this information or the implementation of any of these techniques is a client responsibility and depends upon the client's ability to evaluate and integrate them into the client's operational environment.