Jump to content United States-English
HP.com Home Products and Services Support and Drivers Solutions How to Buy
» Contact HP
More options
HP.com home
Site Preparation Guidehp Integrity Superdome andhp 9000 Superdome > Appendix B Configuration Guidelines

Configuration Guidelines

» 

Technical documentation

Complete book in PDF
» Feedback
Content starts here

 » Table of Contents

 » Index

To achieve the best performance and high availability consider the following factors as shown in order of importance:

Memory Population

Each cell has two memory busses that must be evenly populated to achieve full bandwidth and reduce latency. Configurations of 4 Gbytes, 8 Gbytes, 12 Gbytes, 16 Gbytes, and 32 GBytes per cell accomplish this.

Memory must be symmetric within each partition for interleaving to be even across all the cells, otherwise performance problems can result. Consider the following example:

Example B-1  Incorrect Memory Distribution Across Partitions

If a two cell partition has 2 Gbytes of memory on one cell and 16 Gbytes on another, 4 Gbytes will be interleaved (two from each cell) and 14 Gbytes will be assigned to one cell only. This will result in inconsistent bandwidth and latency problems.

Cell Population

A group of four cells (a quad) resides on one Crossbar Chip (XBC) and has the lowest latency. Quads are slots 0-3, 4-7, 8-11 and 12-15.

Keep partitions with a size less than or equal to four on the same quad.

Memory traffic from one quad to another is routed in pairs. The following pairs share the same path: 0 and 1, 2 and 3, 4 and 5, and 6 and 7.

The Superdome SD32 or hp 9000 Superdome SD32 can have dual links between quads if U-turns are installed on the backplane. To use both links a six-cell partition uses slots 0-3, 5, and 7 as would 4 and 6.

Partitions that share links share common points of failure.

XBC-XBC links can saturate and kill performance. An eight-cell partition on an SD6400 should be spread across cabinets. Cab 0, slots 0-3, 5, 7 and Cab 1, slots 4 and 6.

Partition Size

For best performance, partitions should be either single cell or greater powers of 2: 2, 4, 8, and 16 cells.

Memory interleaving is performed by taking either four or six physical address bits and using them to index into a cell mapping table entry consisting of 16 or 64 sub-entries. For a single cell partition, each sub-entry is loaded with only one cell. A two-cell partition alternates between the two cells. There is one 64 sub-entry cell map entry and a 48-16 sub-entry cell entry.

Partitions that are not powers of two will undergo revisiting. Consecutive accesses will go through all the cell boards and come back through revisiting as many cells as necessary until all memory is utilized and the 64-entry table is filled.

Table 1, SD64 Partition Configurations, shows configurations that yield the best performance. However, memory may be left over. This memory is interleaved in the 16-entry tables and does not always yield the best performance. Depending on the memory configuration and the partition size the memory system performance will always increase, although not always linearly.

Use the following guidelines when building partitions:

  • Build the largest partition first.

  • Keep partitions with four cells or less on the same quad. (Slots 0-3, 4-7, 8-11, and 12-15)

  • Keep the amount of memory on each cell in a partition as equal as possible.

  • Fill empty cabinets before adding to a partially populated cabinet.

  • For best performance, partitions should be either single cell or greater powers of 2: 2, 4, 8, and 16 cells.

  • Always check a new partition for bottlenecks with the formula Qx*Qy/Qt/I. See “XBC-XBC Link Load Equation” for more information.

The partitions in the following table provide good but not always optimum link loading while allowing for growth.

Table B-1 SD64 Partition Configurations

#Cells
per
partition
Cabinet 0 -Cell SlotsCabinet 1 - Cell Slots#Cells
per partition
 0123456789101112131415 
1AIEMCKGOBJFNDLHP16
2AEAECGCGBFBFDHDH8
3AAA CCC BBB DDD 4
4AAAACCCCBBBBDDDD4
5AAAA   ABBBB   B2
6AAAA A ABBBB B B2
7AAAA ABABBBBAB B2
8AAAABABABBBBABAB2
9AAAAAA A    A A 1
10AAAAAAAA    A A 1
11AAAAAAAA    AAA 1
12AAAAAAAA    AAAA1
13AAAAAAAA   AAAAA1
14AAAAAAAA A AAAAA1
15AAAAAAAA AAAAAAA1
16AAAAAAAAAAAAAAAA1

48

 

Table 2 and Figure B-1 “SD64 ” show the SD64 partition configurations.

Table B-2 SD64 Partition Configurations

#Cells
per
partition
Cabinet 0 - Cell Slots#Cells
per
partition
          
1012345578
2AECGBFDH4
3ACACBDBD2
4AAA CCCC2
5AAAABBBB1
6AAAA A A1
7AAAA AAA1
8AAAAAAAA1

 

Figure B-1 SD64

SD64

XBC-XBC Link Load Equation

A simple calculation can be made to evaluate whether or not a particular configuration will have bottlenecks.

  1. The number of cells on one quad, Qx, that talk to another quad Qy

  2. The total number of cells is Qt

  3. The number of links between the two quads is L. (Always 1)

  4. Link load is Qx*Qy/Qt/L. (Link load = 4*4/8/2 = 1).

    Strive always for a link load of 1.0 or less.

The lower the link load, the better off the system. If link loads begin to approach 2, bottlenecks may occur.

For a dual cabinet system, there are six equations covering each link:

  1. Quad 0 talking to Quad 1 L1 = Q0*Q1/Qt/1

  2. Quad 0 talking to Quad 2 L2 = Q0*Q2/Qt/1

  3. Quad 0 talking to Quad 3 L3 = Q0*Q3/Qt/1

  4. Quad 1 talking to Quad 2 L4 = Q1*Q2/Qt/1

  5. Quad 1 talking to Quad 3 L5 = Q1*Q3/Qt/1

  6. Quad 2 talking to Quad 3 L6 = Q2*Q3/Qt/1

I/O Chassis Allocation

For best high availability, I/O devices are connected to different I/O Bays to reduce single points of failure.

Rules dictate that the roots are selected first, then all other I/O are added.

I/O can connect anywhere, but tools will guide you to the best selection.

Printable version
Privacy statement Using this site means you accept its terms Feedback to webmaster
© Hewlett-Packard Development Company, L.P.