Teach you to choose server memory from technical details

If IT people don't know much about the server memory market, choosing the right server memory is not an easy task.

Virtualization increases the number of workloads that run on each server, but more and more computing requirements require IT staff to have better hardware. Memory is already a critical resource for the server, and in general, the virtual server consumes memory before it consumes other computing resources, such as the processor core or clock cycles. This article outlines the key attributes of several major memory types on the market today and helps IT professionals match different levels and performances of memory to server and expected workload requirements.

Server Memory Basic Properties

In addition to the underlying 64-bit or 72-bit data bandwidth and speed identifiers, today's memory modules are categorized using a range of terms, as IT administrators or technicians, Familiarize yourself with these terms to make it easier to understand server configuration and discuss the effectiveness of memory attributes with server vendors.

Today's server memory is typically categorized using the transfer rate in millions of transfers per second (MT/s). A typical dual in-line memory module (DIMM) device provides a transmission rate of 1,066 MT/s, 1,333 MT/s, 1,600 MT/s or 1,866 MT/s. Keep in mind that this does not represent the data transfer rate, it only means that this device can run these number of transfer cycles. If you want to find the data transfer rate – bytes (B) per second – directly multiply the transfer rate by 8. For example, a leading edge DIMM with a transfer rate of 2,133 MT/s will provide a data transfer rate of 17,066 MB per second, which is equivalent to approximately 17 Gbps of bandwidth.

A "memory rank" is a set of dynamic RAM (DRAM) memory chips connected to the same chip. The computer is then used at the same time via a 64-bit data bus or a 72-bit data bus using error correction code (ECC). The actual physical memory chip can vary in a rank. For example, an x16 (16-bit) rank memory chip can use four 16-bit DRAM memory chips to create a 64-bit rank. In contrast, an x8 (8-bit) rank can use a total of eight 8-bit DRAM memory chips to complete a 64-bit rank; if the memory bank supports ECC, more memory chips can be added. A memory module such as a DIMM can include 1, 2, 4 or 8 ranks called "single," "dual," "quad" and "octal". Single rank DIMMs are the cheapest and offer only the lowest amount of memory.

Ranks can be confusing, but they are a key part of memory geometry. For example, when you have a DIMM and have 8 chips on each side – a total of 16 DRAM chips are on the DIMM, each of the 16 chips represents a rank, and each rank has a bandwidth of 8 bits. In fact, this DIMM will end with 2 ranks, each of which is a 64-bit bandwidth, located on either side of the memory bank.

The server memory controller's support for the DIMMs that need to be installed is limited by the number of ranks and also varies with the total capacity used by the DIMMs. For example, a system may support the use of ranks of four low-density DRAM chips, resulting in DIMMs for high-density DRAM chips that only bring less rank to the system.

A "memory Channel", sometimes called “bank” represents a group of ranks. All ranks added to the channel can be a single DIMM, or a few DIMMs. A typical server usually supports a large number of channels and also allows a large number of memory modules to be installed. In most cases, each DIMM in a channel must be logically identical. The important point is that when the channel needs to be filled, the system documentation needs to be reviewed to determine all the constraints or system requirements.

Basic Memory Types

Memory modules are now planned by type: divided into registration, unbuffered, and load reduction. Each type provides some trade-offs because the types cannot be mixed within the same server, so it's important to consider the best combination of performance, reliability, efficiency, and cost before you make a decision.

Registering DIMMs (RDIMMs) works by buffering DRAM addresses, control and clock signals. Buffering improves signal strength, simplifies electronic load problems, and provides a good combination of speed (output), capacity and rank configuration so RDIMMs are the most popular models. Singal or Dual rank RDIMMs can support 1,600 MT/s output with 16 DIMMs per DIMM. The 32GB RDIMMs support quad rank configuration (that is, with more chips on each DIMM), but this usually limits the number of DIMMs installed on each channel and reduces the transfer speed to 1,066 MT/s. In summary, RDIMMs will be the best choice when DIMM reliability depends on ECC or when the server needs to support large amounts of memory.

In contrast, unbuffered DIMMs (UDIMMs) do not buffer address, control, and clock signals. Because the buffered electron delay is removed, the lack of buffering can increase a small percentage of DIMM performance. However, UDIMMs impose a larger electronic load on the host server, limiting the maximum output to less than 1,600 MT/s, reducing the DIMM capacity to 4 GB, and limiting only 2 DIMMs per channel. If you really want to use it, UDIMMs can be used on alternate servers that don't require a lot of memory, and may benefit from a slightly lesser memory latency.

Load-reducing DIMMs are similar to registered DIMMs and have caches, but LRDIMMs use a buffering method to significantly reduce the electronic load on each DIMM. This allows up to 3 large-capacity 32-bit quad rank DIMMs to run at a transfer rate of 1,333 MT/s on a single channel. At the same time, the server can achieve the highest total memory capacity, but if more than 3 DIMMs are used on each channel, the transfer rate will drop to 1,066 MT/s. LRDIMMs are still being adopted by some server vendors, so if you want to use them, make sure your designated server supports LRDIMMs. For example, HP Prolian G8 servers support LRDIMMs, while older versions of G7 and G6 do not.

Memory is a critical resource for any virtualized data center, so choosing a memory bank can have a huge impact on the memory capacity, reliability, and performance used by each workload in the system. Ok, you have learned some basic concepts of contemporary memory naming, and you can make better choices for your server memory configuration to maximize the IT hardware budget you use.