Advertisement
QuickSpecs
Configuration
I/O
I/O
I/O
I/O
14
15
16
17
18
19
20
21
Performance
Performance
Performance
Performance
22
23
24
25
26
27
28
29
30
31
32
33
34
35
Single System
Single System
Single System
Single System
36
HP Integrity Superdome Servers: 16-socket, 32-
HP Integrity Superdome Servers: 16-socket, 32-
HP Integrity Superdome Servers: 16-socket, 32-
HP Integrity Superdome Servers: 16-socket, 32-
One cell in every partition must be connected to an I/O chassis that contains a Core I/O card, a card connected
to boot media, a card connected to removable media, and a network card with a connected network.
A partition cannot have more I/O chassis than it has active cells.
Removable media device controller should be in slot 8 of the I/O chassis.
Core I/O card must be in slot 0 of the I/O chassis.
Boot device controller should be in slot 1 of the I/O chassis
PCI-X high bandwidth I/O cards should be in the high bandwidth slots in the I/O chassis
Every I/O card in an I/O chassis must be assigned to a valid physical location.
Every I/O chassis in a Superdome complex must be assigned to a valid physical location
The amount of memory on a cell should be evenly divisible by 4 GB if using 512-MB DIMMs or 8 GB if using 1-
GB DIMMs, i.e. 8, 16 or 32 DIMMs. The cell has four memory subsystems and each subsystem should have an
echelon (2 DIMMs) populated. The loading order of the DIMMs alternates among the four subsystems. This rule
provides maximum memory bandwidth on the cell, by equally populating all four memory subsystems.
All cells in a partition should have the same number of processors.
The number of active CPUs per cell should be balanced across the partition, however minor differences are OK.
(Example: 4 active CPUs on one cell and three active CPUs on the second cell)
If memory is going to be configured as fully interleaved, all cells in a partition should have the same amount of
memory (symmetric memory loading). Asymmetrically distributed memory affects the interleaving of cache lines
across the cells. Asymmetrically distributed memory can create memory regions that are non optimally
interleaved. Applications whose memory pages land in memory interleaved across just one cell can see up to 16
times less bandwidth than ones whose pages are interleaved across all cells.
If a partition contains 4 or fewer cells, all the cells should be linked to the same crossbar (quad) in order to
eliminate bottlenecks and the sharing of crossbar bandwidth with other partitions. In each Superdome cabinet,
slots 0, 1, 2 and 3 link to the same crossbar and slots 4, 5, 6 and 7 link to the same crossbar.
A Core I/O card should not be selected as the main network interface to a partition. A Core I/O card is a PCI X
1X card that possibly produces lower performance than a comparable PCI X 2X card.
The number of cells in a partition should be a power of two, i.e., 2, 4, 8, or 16.
Optimal interleaving of memory across cells requires that the number of cells be a power of two. Building a
partition that does not meet this requirement can create memory regions that are non optimally interleaved.
Applications whose memory pages land in the memory that is interleaved across just one cell can experience up
to 16 times less bandwidth than pages which are interleaved across all 16 cells.
Before consolidating partitions in a Superdome 32-socket or 64-socket system, the following link load
calculation should be performed for each link between crossbars in the proposed partition.
Links loads less then 1 are best. As the link load begins to approach 2 performance bottlenecks may occur.
For crossbars X and Y
Link Load = Qx * Qy / Qt / L, where
- Qx is the number of cells connected to crossbar X (quad)
- Qy is the number of cells connected to crossbar Y (quad)
- Qt is the total number of cells in the partition
- L is the number of links between crossbar X and Y (2 for Superdome 32-socket systems and 1 for Superdome 64-
socket systems)
Maximum performance for optimal configurations (power of two cells, uniform memory across cells, power of two
DIMM ranks per cell)
(If rule #30 cannot be met, rule #31 is recommended) Non-power of two cells, but still uniform memory across
cells, power of two DIMM ranks per cell, uniform type of DIMM.
(If rule #30 or #31 cannot be met, rule #32 is recommended) Same amount of memory in each cell, but
possibly different memory types in each cell (for instance, a two cell configuration with 8 512MB DIMMs in one
cell, and 4 1GB DIMMs in the other). Differences in memory across different cells within the same partition
should be minimal for the best performance.
Same amount of memory in each cell, but non optimal and/or mixed loading within a cell (for instance, a two
cell configuration with 16 512MB DIMMs and 8 1GB DIMMs in each cell).
Non-uniform amount of memory across cells (this needs to boot and run, but performance is whatever you get).
For the same amount of total memory, best performance is with a larger number of smaller size DIMMs.
Each cell should have at least two active CPUs.
DA - 11717
North America — Version 15 — January 3, 2005
socket, and 64-socket
socket, and 64-socket
socket, and 64-socket
socket, and 64-socket
Page 27
Advertisement
