The last of the three big announcements that alter the way large IBM systems will be used in the future is IBM’s belated recognition that Amdahl Corp was onto rather a good thing with its Multiple Domain Feature hardware partitioning facility. The facility was first introduced by IBM last summer under VM/XA, under the name Multiple High Performance Guest Support. The concept there was to extend to several virtual machines the facility previously available to only one guest under VM, that of hogging all the resources of the machine. That in itself was a feature originated by Amdahl Corp back in about 1980, and picked up by IBM a couple of years later. Now however, as with the Amdahl offering, users can forget about VM. The new facility, based on Multiple High Performance Guest Support, but renamed Processor Resource/Systems Manager because without VM there are no guests, provides flexible partitioning of the IBM 3090 Processor Complex Enhanced models – note: E models only again into a maximum of four logical partitions. A maximum of eight logical partitions can be created on 3090 multiprocessing models that could always be partitioned into two separate systems. In other words the 3090-280E, 400E, 500E and 600E – and they have to be running in dual image mode, not single image. The new PR/SM feature – IBM wants you to call it Prism – supports the newly announced Enterprise Systems Architecture/370 and will be available in the US in August. One justification for offering the new facility is that it will provide a fairly convenient environment for testing the new Enterprise Systems Architecture release of MVS-SP without unac.cw 8 ceptably inefficient use of processor resources. Event-driven The 3090 PR/SM feature enables event-driven, logical partitioning function of the Processor Complex and machines with it fitted supports three environments – System/370, ESA/370 (which supports both the 370-XA and ESA/370 operating systems), and the Logically Partitioned mode. When this mode is chosen, the operator can define the resources that are to be allocated to each logical partition and most resources can be reconfigured without requiring a power-on-reset. Once a logical partition is activated, a supported operating system can be booted into that logical partition. Central Storage and optional Expanded Storage are defined to logical partitions prior to partition activation. When a logical partition is activated, the storage resources are allocated in 1Mb contiguous blocks. These allocations are static upon activation. Sharing of allocated Central Storage or Expanded Storage between multiple logical partitions is not allowed. Individual channel paths are also defined to a logical partition and a channel path can only be allocated to one logical partition at a time, although a device can be shared between logical partitions by using a separate channel path from each partition. Channel paths can also be dynamically reconfigured between logical partitions. Individual central processors can be dedicated to a single logical partition or shared between multiple logical partitions and the allocation of processors to a logical partition is made when the partition is activated; use of processor resources shared between logical partitions can be modified by operator commands while the logical partitions are active. If there is a Vector Facility installed on a central processor, it is available to all partitions that will execute on that processor. PR/SM includes four additional byte multiplexer capable channels per processor side for a total of eight per side. The PR/SM feature also includes enhancements to the System Activity Display so that the operator can see what is going on. Resource Management Facility now supports the PR/SM feature and can report processor usage and partition data for each configured logical partition, and input-output performance for the entire system. Unexpected One of the unexpected features of the partitioning process is that it can actually improve aggregate performance over a non partitioned machine – no doubt b
ecause of those extra channels. In general, says IBM, for interactive workloads such as IMS/VS or TSO/E under MVS/XA, or CMS under VM/SP HPO, the aggregate performance, as measured by Internal Throughput Rate, of multiple logical partitions with dedicated processors is projected to be in the range of 88% to 110% of the throughput rate of a single system operating in S/370 mode or ESA/370 mode on the total configuration. When large single image processing is not required, increased throughput is possible due to the reduced functional requirement. As an example, IBM says that prototype measurements of a laboratory MVS/SP 2.2 IMS workload on a 3090 Model 200E configured as two equal dedicated uniprocessor logical partitions achieved an throughput rate 103% of that of the same workload operating in 370-XA mode on the entire dyadic. A similar prototype measurement of a laboratory VM/SP HPO 5 CMS workload indicated that a 3090 Model 200E logically partitioned as two dedicated uniprocessors can achieve a throughput that is 106% of that of the same workload operating in S/370 mode on the entire dyadic. Multiple logical partitions that share, rather than dedicate, processors will require more central processing unit usage, but the ability of PR/SM to recognise unused processor resources and balance workloads may result in an improved External Throughput Rate with non-dedicated processors. In particular, says IBM, for workloads that exhibit fluctuations in processing demands, the consolidation of several systems as non-dedicated partitions of a processor complex may improve responsiveness and increase external throughput. The implementation of the PR/SM feature uses 3090E hardware mechanisms that are also used to implement the Start Interpretive Execution facility when running the 3090E models in ESA/370 mode so that if this facility is used in a logical partition, SIE performance will be reduced relative to operation in ESA/370 mode. The performance reduction depends on the frequency of SIE invocation. The PR/SM feature, with Logically Partitioned operation, is set for August; no prices.
