System p

from Wikipedia, the free encyclopedia

System p (formerly pSeries ) is a RISC -based Unix computer family from IBM . The pSeries was the successor to the RS / 6000 series (RISC System / 6000) introduced by IBM in 1990 .

The p in the name stands for performance , not, as variously assumed, for Power PC .

CPUs

The pSeries is based on the Power -CPU, the PowerPC is used in smaller models . Up to 32 of these power CPUs can work together in a pSeries computer. IBM delivers the proprietary AIX or Linux operating systems for the pSeries .

pSeries computers support (from the p630 upwards) the dynamic logical partitioning LPAR . The (2003) largest pSeries computer p690 can be divided into up to 32 independent partitions and manages 32 CPUs and 512 GB main memory. The resources can be dynamically moved from one LPAR to another, since AIX 5.2 no reboot is necessary.

history

In 2004, IBM eServer p5 (see below) and OpenPower -Server were introduced as successors to pSeries.

In 2006 the pSeries was rebranded to the name "System p" and new models were introduced in the process. These p5- {505, 510, 520, 550, 560, 570, 575, 590 and 595} work with 1-64 POWER5 CPUs. An "A" in the model designation indicates POWER5 + processors, a "Q" quad-core CPUs.

The Power5-based platforms System i and System p have been physically practically identical since the announcement of i5 and p5. The difference is made by the characteristics of the chosen operating system OS / 400 (now i5 OS), AIX or Linux .

IBM eServer p5

IBM Power5 p595 front
IBM Power5 p595 front opened
IBM System p5 logo

The IBM eServer p5 uses POWER5 CPUs that support micro-partitioning and SMT . The IBM eServer p5 server range extends from a 1-2-way p5 505 server to a 64-way p595. p5 servers can be operated with AIX and Linux and can handle logical partitioning, i. H. the installation of several operating system instances on one server. The operating systems in the so-called partitions ( LPAR - Logical Partition) are independent of each other, and although they run on the same server, they only have access to the hardware assigned to them. There is a special series of p5 servers: OpenPower 710 (1-2 ways) and OpenPower 720 (1-4 ways), servers that are exclusively intended for the use of Linux and do not allow the operation of AIX . All p5 and OpenPower servers are based on POWER5 technology and can run 32-bit and 64-bit programs in parallel and natively.

Virtualization on IBM eServer p5 and OpenPower

All servers based on p5 are capable of logical partitioning, dynamic logical partitioning and micro-partitioning. Partitioning is the distribution of hardware resources to individual operating system instances, dynamic denotes the ability to move these resources without restarting and micro- partitioning denotes the ability to allocate fractions of the processors to the individual partitions. The virtualization technology comes from the mainframe area and is used today in various software and hardware solutions.

With p5 servers, the virtualization is carried out by the processor technology supported in the firmware , called hypervisor . The hypervisor divides the hardware of the system and makes parts of it available to the operating system in a logical partition ( LPAR ). Each LPAR thus receives a “local” firmware (a kind of BIOS) that only makes the hardware visible to it that has been assigned to it.

A CPU can be distributed to up to 10 LPARs, i.e. H. Up to 40 independent operating system instances can be installed in logical partitions on a 4-way machine. On the currently (2005) largest p5 server, the 64-way p595, 254 operating system instances are possible. Both AIX (5.2 and 5.3) and Linux can be installed on these at the same time .

In contrast to most software-based virtualization solutions, there are hardly any restrictions on the scalability of a single partition with the POWER architecture. The partition can contain and use from 0.1 up to 64 processors. Dedicated and virtualized resources can be mixed within an LPAR and a server. The overhead resulting from virtualization is difficult to prove because the hypervisor layer is always present on the p5 servers, i. H. all benchmarks were also made with hypervisor.

Thanks to virtualization and automatic load distribution, a much higher overall system utilization is achieved. The virtualization of the I / O resources also saves adapters .

requirements

A special computer called HMC (Hardware Management Console) is required for the virtualization device. This can manage up to 32 servers and / or up to 128 logical partitions (LPARs) and, in addition to the main task of dividing the machines into LPARs, also takes on other functions, for example the HMC replaces the serial connections to the individual partitions and thus makes additional ones No need for consoles or KVM switches . Operation WITHOUT HMC can also be virtualized under POWER5 + for small to medium-sized systems. In this case, the functions of the HMC (with certain functional restrictions) are taken over by the so-called IVM (Integrated Virtualization Manager).

HMC is the graphical interface for classifying machines and for changing resource allocation. The HMC is not necessary for the operation of the partitioned systems themselves and can be switched off or replaced at any time. With the Linux-only OpenPower systems, activation is required for the virtualization, which is subject to a fee.

CPU virtualization

The CPU can be virtualized in two ways: CPUs can be allocated to a partition (operating system instance), or parts of the computing capacity are allocated to a partition (shared). Both types of allocation can be mixed on one server. Partitions with dedicated CPUs do not participate in the automatic load balancing. All CPUs that are not allocated on a dedicated basis remain in a so-called "shared CPU pool". This pool is used proportionally by all LPARs that are set up in "shared" mode. If the assigned cycles are not used by an LPAR, they are returned to the pool and can be used by the instances that currently require more CPU power.

Storage virtualization

The main memory is allocated to individual LPARs in steps of a size that can be set for the entire system (memory region size). The memory accesses are implemented by the hypervisor in such a way that no partition can access the memory areas of the other. AIX (from 5.2) can handle dynamic (= during operation) memory allocations. Linux must be restarted after changing the memory allocation.

Network virtualization

The hypervisor functions as a virtual network switch . The virtual switch is VLAN- capable. Virtual network cards can be assigned to all partitions ; from the operating system's point of view, these behave largely like real physical cards.

Physical and virtual network cards can be assigned at the same time. For example, one of the partitions that has both a physical and a virtual network card can act as a router or bridge for the other partitions that are only equipped with virtual network cards. This task is usually taken over by the VIO server (see hard disk virtualization). A physical and a virtual network card of the VIO server are configured as a bridge (AIX terminology: SEA - Shared Ethernet Adapter). If the virtual network cards of the other LPARs are in the same network and VLAN, they can be accessed transparently from the outside (without specifying a gateway ).

Hard disk virtualization

The hypervisor is only able to grant access to a PCI-X slot of a partition; individual hard drives cannot be distributed. A special partition is required for this, called a VIO server (virtual I / O server). This is assigned the PCI-X slot, in which the adapter is located, to which the hard disks are connected, and it is able to make parts of these disks available to the other LPARs as entire virtual hard disks. There is a version of the VIO server that is based on AIX , but the same functionality can also be achieved with Linux .

On the client LPAR side (those that use virtual hard disks that have been released by the VIO), only a driver for a virtual SCSI adapter is required. This behaves in exactly the same way as a normal SCSI driver .

See also

Web links

To the IBM eServer p5 :

Individual evidence

  1. For the use of Linux on IBM pSeries see the following Redbook ( Memento of February 6, 2004 in the Internet Archive ).