Shulou(Shulou.com)06/02 Report--

When considering deploying a super-converged architecture, many enterprises tend to use flexible hardware configurations rather than just one brand. SmartX super-converged solution provides two delivery modes: Halo all-in-one machine and SMTX OS pure software, which can better meet the needs of different users. For users who choose SMTX OS software solutions, SmartX has specially launched this hardware configuration guide-"SMTX OS Super Convergence hardware selection recommendation Guide" (hereinafter referred to as "Guide"), which aims to provide reference from a professional point of view to help users choose their own hardware products more specifically and make super-convergence hardware configuration easier.

SmartX divides the Guide into three parts: stand-alone hardware configuration, hardware resource reservation scheme, and expansion selection. In the first issue, we extracted the CPU chapter from the overview and stand-alone hardware configuration (click Review). Today, we will continue to present to you the second issue-- memory of stand-alone hardware configuration.

If you want to get the complete Guide in advance, you can click here to enter the download channel.

Memory of stand-alone hardware configuration

When users do the overall super-convergence application planning, they can first plan the total memory capacity requirements according to the actual application requirements, and then make memory configuration choices according to the overall memory capacity requirements. SMTX OS currently does not specify the brand of memory, the mainstream brand is fine, but it is best to ensure that the memory configured by a single server is of the same brand and model.

1. Recommended amount of memory

Intel ®Xeon ®Scalable Processor supports six memory channels Channel per CPU, and a maximum of two memory DIMM per channel. Two-way rack servers can support up to 12 memory channels and 24 memory. When the amount of memory is 12:00, memory will take up all the Channel, that is, one memory for each channel. When the amount of memory is 24:00, memory accounts for all Channel and DIMM. In these two cases, the overall memory performance is the best. Optimal memory performance means higher memory bandwidth, so it is recommended to configure 12 or 24 bars of memory.

If you cannot guarantee the amount of 12 or 24 memory, you should also ensure that the amount of memory can be divided by 4, or the number of memory in a single CPU configuration can be divided by 2. While ensuring the symmetrical configuration of two CPU memory, it is also necessary to ensure the symmetrical configuration of a single CPU memory.

The following is a specific description.

The schematic diagram of the Intel ®Xeon ®Scalable Processor CPU memory layout is as follows:

Schematic diagram of Intel ®Xeon ®Scalable Processor CPU memory layout

The memory configuration layout of a single CPU is represented by the number of memories per channel: "A:B:C", "D:E:F". For example, the configuration of 2 D:E:F 1 is marked in orange in the figure below:

For example, the configuration of 2RV 1V 1R 1D 2R 1R 1R 1

The configuration of the memory stick should conform to three Balanced principles as far as possible:

The memory configuration is consistent under all channels (channels with memory configured).

The memory configurations of two memory controllers under the same CPU are consistent.

The memory configuration of all CPU under the same server is consistent.

The relative performance achieved by the number of memory per CPU configuration (100% bandwidth value tested by Stream in a 12-DIMM configuration) is compared as follows:

Source: https://lenovopress.com/lp0742-intel-xeon-scalable-family-balanced-memory-configurations

The comparison diagram is as follows:

Relative performance under different memory configurations

Therefore, we can see that when the number of memory in a single CPU is 12 or 6, the memory performance is optimal, that is, the number of memory in the configuration of the two CPU is 24 or 12, but there is little difference between the two configurations. From the point of view of subsequent memory expansion of a single server, it is recommended to configure 12 memory for a single server to reserve memory slots for subsequent expansion.

two。 Recommended memory capacity per bar

At present, the main memory capacity is divided into 16 G and 32 G and 8 G memory is basically eliminated. 128 GB of memory is expensive and configuration is not recommended when large memory capacity is not needed. The market price of a single 32-gigabyte memory is generally better than that of two 16-gigabyte memory, and 32-gigabyte memory is preferred. If a single server needs to configure a large amount of memory, you can also consider 64 GB of memory, but compare the actual price difference with 32 GB of memory.

3. Total memory capacity configuration recommendation

The matching relationship between memory quantity and capacity is selected according to user demand, performance, price and capacity expansion. If the overall memory capacity requirement is 384 GB, it is recommended to configure 12 strips of 32 G memory; the overall memory capacity requirement is 192 G, and it is recommended to configure 12 strips of 16 G memory.

If 24 or 12 pieces of memory cannot be met, it is recommended to configure a certain amount of 32 G memory. If the required capacity is 128 G, 4 bars of 32 G memory should be configured; if the required capacity is 512 G, 16 bars of 32 G memory should be configured. Or configure flexibly according to the table in 1 to ensure maximum memory performance.

4. Recommended memory type

At present, the mainstream server memory for RDIMM and LRDIMM,UDIMM memory has been basically eliminated. The main differences between the three types of memory are as follows:

UDIMM: also known as Unbuffered DIMM. When data is transferred from CPU to each memory particle, UDIMM requires that the data transmission distance between CPU and each memory particle is equal, so that parallel transmission can be effective. This requires a very high manufacturing process, and it is difficult to achieve high density and high frequency, so the capacity and frequency of UDIMM are low.

RDIMM: also known as Registered DIMM. In order to ensure the effectiveness of parallel transmission, RDIMM adds a register to the memory bar for forwarding. It is located between the CPU and memory particles, thus reducing the distance of parallel transmission. At the same time, because of the high efficiency of registers, the density and frequency of RDIMM are easy to increase.

LRDIMM: also known as Load Reduced DIMM. When the server is configured with RDIMM, the memory bus is in parallel mode, and all DRAM is controlled by the processor's memory controller. As more and more DRAM are integrated into the RDIMM, the power load of memory modules increases (this is called memory column-memory modules have single-column, double-column, and four-column options).

As more columns are installed in the memory channel, memory speed is reduced and / or the use of additional memory slots is limited. LRDIMM uses memory buffering chips to get rid of these limitations and further increase memory support capacity accordingly. When the server is equipped with only LRDIMM, the memory controller in the processor automatically switches to serial mode-all data, commands, and control signals are packaged and transferred to the memory buffer on the LRDIMM, which then handles all read and write operations to the DRAM chip.

Therefore, LRDIMM can achieve a larger single memory capacity than RDIMM, but due to the use of buffer chips instead of registers, the latency will also increase.

To sum up, from the current point of view, LRDIMM is more suitable for large overall memory capacity configuration, while when the overall memory capacity is not too high, RDIMM can provide better performance when the overall memory capacity is the same.

5. Memory frequency recommendation

At present, the mainstream frequencies of memory are 2666 MHz and 2400 MHz, and the maximum memory frequencies supported by different series of CPU are also different. However, due to the small difference in memory prices between the two frequencies, and considering the consistency of the overall components of the customer data center, 2666 MHz memory is recommended for memory configuration. For CPU that cannot support up to 2666 MHz, memory can be reduced to the frequency supported by CPU.

To sum up, the overall memory recommendations are as follows:

Amount of memory: it is recommended to configure 12 items to provide better memory bandwidth and reserve space for subsequent capacity expansion. (if 12 pieces of memory cannot be matched, the amount of memory can be selected according to the overall memory capacity requirements.)

Single strip capacity: 32 GB memory per strip is recommended, and the cost is better than 16 GB memory per strip.

Memory type: RDIMM memory is recommended.

Memory frequency: 2666 MHz memory is recommended.

Learn more about SmartX superfusion products: https://www.smartx.com

Download the complete Guide: https://www.smartx.com/smtx-os#doc-smtx-os-recommend

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.