Fujitsu/Fujitsu Siemens Computers PRIMEPOWER Enterprise Servers
© 2003 Gartner, Inc. and/or its Affiliates. All Rights Reserved.
DPRO-97941
17 April 2003
4
which also supported up to 128 processors, but required additional cabinets for configurations of more
than 64 CPUs and for I/O extension.
The PRIMEPOWER 2500’s system boards do not include any internal PCI slots, but they each connect
up to two PCI/disk boxes. This means a 64-way version can support up to 16 PCI/disk boxes, and a 128-
way server can support up to 32 PCI/disk boxes. The 64-way supports up to 256GB of memory, and the
128-way supports up to 512GB. The 2500 has two redundant Level-2 crossbars connecting the two
backplanes to which the system boards are attached. If one crossbar fails, the system will isolate the
faulty components and automatically restart with a reduced configuration (and half the bandwidth). Each
backplane is clocked at 520MHz and has a sustained throughput of 66.5 GB/second giving 133
GB/second throughput across the Level-2 crossbars in total. The 2500 also has dual redundant system
clocks.
RAS Features
The 2500 has the highest levels of availability features in the PRIMEPOWER product line, but there are
extensive levels of error checking and automatic error correction built into the SPARC64 V processor. As
well as error checking and correcting (ECC) memory, all three XA models also have ECC checking on the
data path and on the PCI bridges.
The SPARC64 V has some unique error-correcting features for a reduced instruction set computer (RISC)
processor. There is parity checking at the register and execution unit level. If an intermittent parity error
occurs, there is automatic instruction retry by the hardware. If the retries are unsuccessful and parity
errors continue to appear, the system will automatically “degrade” the processor configuration, removing
the faulty processor from the active configuration. The Level-2 (L2) cache has ECC protection (double-bit
error detection and single-bit error correction) for both data and tags. The Level-1 (L1) caches have ECC
protection for data, while L1 tags are duplicated and immediately corrected if a parity error is detected. All
this is handled automatically by the hardware without the intervention of the operating system (OS).
Hardware error logs are collected during operation, however, and reported to the System Control Facility
(SCF). There are over 90,000 monitors on a 32-way PRIMEPOWER model 2500 when the system is
running. Any error events are reported automatically via the Call-home feature.
The Solaris operating system supports dynamic reconfiguration, so that a system board can be logically
detached from a partition and powered down—partitions can only be powered down at the system board
or PPAR level, however, not at the XPAR level (see below). Likewise, PCI boards that are part of the
system board on the 900 and 1500 can only be logically detached and physically replaced as part of the
whole system board. Individual PCI adapters can be hot-plugged, however, and Solaris supports I/O
multipathing to allow redundant I/O connections.
XPAR
The PRIMEPOWER XA models feature two types of hard partitions. PPARs are the more usual form of
hard partitioning created at the system board boundaries. They are similar to Sun’s Dynamic System
Domains and HP’s nPars (hard partitions). The PRIMEPOWERs’ PPARs consist of complete eight-way
system boards. XPARs are also hard partitions, but they allow more granularity than PPARs. Switching on
the system board allows it to be divided up into two or four XPAR hard partitions. The most granular
XPAR on the 900 and 1500 can thus consist of two CPUs and 1GB of memory. This means that hard
partitions on the PRIMEPOWER XA models can not only span complete system boards, but can also
include part system boards (see figure). Clients should note that on the PRIMEPOWER 2500, granularity
is limited to four CPUs.
Figure 1: PRIMEPOWER XPARs and PPARs
