IBM eX5 Implementation Guide

24 IBM eX5 Implementation Guide

Figure 2-6 Relative memory performance based on DIMM placement (one processor and two memory cards shown)

2.3.3 Memory ranking

The underlying speed of the memory as measured in MHz is not sensitive to memory

population. (In Intel Xeon 5500 processor-based systems, such as the x3650 M2, if rules

regarding optimal memory population are not followed, the system BIOS clocks the memory

subsystem down to a slower speed. This situation is not the case with the x3850 X5.)

Unlike Intel 5500 processor-based systems, more ranks are

better for performance in the

x3850 X5. Therefore, quad-rank memory is better than dual-rank memory, and dual-rank

memory is better than single-rank memory. Again, the frequency of the memory as measured

in MHz does not change depending on the number of ranks used. (Intel 5500-based systems,

such as the x3650 M2, are sensitive to the number of ranks installed. Quad-rank memory in

those systems always triggers a stepping down of memory speed as enforced by the BIOS,

which is not the case with the eX5 series.)

Performance test between ranks

With the Xeon 7500 and 6500 processors, having more ranks gives better performance. The

better performance is the result of the addressing scheme. The addressing scheme can

1

Each processor:

2 memory controllers

2 DIMMs per channel

8 DIMMs per MC

Mem Ctrl 1 Mem Ctrl 2

1.0

2

Mem Ctrl 1 Mem Ctrl 2

Each processor:

2 memory controllers

1 DIMM per channel

4 DIMMs per MC

0.94

Mem Ctrl 1

Memory card

DIMMs

Channel

Memory buffer

SMI link

Memory controller

3

Mem Ctrl 1 Mem Ctrl 2

Each processor:

2 memory controllers

2 DIMMs per channel

4 DIMMs per MC

0.61

Relative

performance

4

Mem Ctrl 1 Mem Ctrl 2

Each processor:

2 memory controllers

1 DIMM per channel

2 DIMMs per MC

0.58

5

Mem Ctrl 1 Mem Ctrl 2

Each processor:

1 memory controller

2 DIMMs per channel

8 DIMMs per MC

0.51

6

Mem Ctrl 1 Mem Ctrl 2

Each processor:

1 memory controller

1 DIMM per channel

4 DIMMs per MC

0.47

7

Mem Ctrl 1 Mem Ctrl 2

Each processor:

1 memory controller

2 DIMMs per channel

4 DIMMs per MC

0.31

8

Mem Ctrl 1 Mem Ctrl 2

Each processor:

1 memory controller

1 DIMM per channel

2 DIMMs per MC

0.29

1

0.94

0.61

0.51

0.47

0.31

0.29

0.58

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

12345678

Configuration

Relative memory performance

Memory configurations

Chapter 2. IBM eX5 technology 25

extend the pages across ranks, thereby making the pages effectively larger and therefore

creating more page-hit cycles.

We used three types of memory DIMMs for this analysis:

򐂰 Four GB 4Rx8 (four ranks using x8 DRAM technology)

򐂰 Two GB 2Rx8 (two ranks)

򐂰 One GB 1Rx8 (one rank)

We used the following memory configurations:

򐂰 Fully populated memory:

– Two DIMMs on each memory channel

– Eight DIMMs per memory card

򐂰 Half-populated memory:

– One DIMM on each memory channel

– Four DIMMs per memory card (slots 1, 3, 6, and 8; see Figure 3-16 on page 76)

򐂰 Quarter-populated memory:

– One DIMM on just half of the memory channels

– Two DIMMs per memory card

Although several benchmarks were conducted, this section focuses on the results gathered

using the industry-standard STREAM benchmark, as shown in Figure 2-7.

Figure 2-7 Comparing the performance of memory DIMM configurations using STREAM

Taking the top performance result of 16x 4 GB quad-rank DIMMs as the baseline, we see how

the performance drops to 95% of the top performance with 16x 2 GB dual-rank DIMMs, and

89% of the top performance with 16x 1 GB single-rank DIMMs.

You can see similar effects across the three configurations based on eight DIMMs per

processor and four DIMMs per processor. These results also emphasize the same effect that

is shown in 3.8.3, “Maximizing memory performance” on page 84 for the x3850 X5, where

performance drops away dramatically when all eight memory channels per CPU are not used.

Relative STREAM Triad Throughput

by DIMM population per processor

100

98

55

95

89

52

89

73

42

0 20406080100120

16x 4GB (4R)

8x 4GB (4R)

4x 4GB (4R)

16x 2GB (2R)

8x 2GB (2R)

4x 2GB (2R)

16x 1GB (1R)

8x 1GB (1R)

4x 1GB (1R)

Relative Memory Throughput