【 Weak current College 】 Server detailed server memory

Server memory is the memory (RAM), and ordinary PC (personal computer) machine memory in appearance and structure does not have any obvious differences of substance, mainly in memory on the introduction of new unique technologies, such as ECC ChipKill, hot-swappable, technology, etc., with very high stability and error-correcting performance.

Server memory technology:

(1) ECC

In a normal memory, often use a technique that Parity check codes with bits (Paritycheckcodes) is widely used in multi-monitor code (errordetectioncodes), they increase a check bits to each information of characters (or bytes) and can detect a character in all odd (occasionally) and bit error, but there is a drawback, Parity when computer found there is an error for a Byte, and you cannot determine which bits in error, that error cannot be corrected. Based on the above, generated a new memory error correction technology, ECC, ECC itself is not a memory model that is not a memory dedicated technology, it is a widely used in various areas of computer instructions, is a kind of instruction error correction technology. ECC full name in English is "ErrorCheckingandCorrecting", the corresponding name is called "error checking and correction", from the name we can see that its main function is to "find and correct the error," it is better than parity calibration technology more advanced aspects of the main is that it not only can find the error, and correct these errors, the error correction after computer to properly perform the following tasks to ensure server uptime. Say it is not a memory model, it is because it is not a new affected memory structure and storage speed of technology, it can be applied to different memory types, like the former to "odd correction" memory, it is not a memory, the start of the application of this technology is EDO memory, SD and application, and ECC memory mainly from SD memory began to be widely applied, and the new DDR, RDRAM also have a corresponding application, the current mainstream ECC memory is actually a kind of SD memory.

(2) Chipkill

IBM Chipkill technology is designed to address the current server memory ECC technology developed by inadequate, is a new ECC memory protection standards. We know that the ECC memory can also detect and correct single-bit errors, but if the simultaneous detection of two more bits of data has errors, the General could do nothing about it. The reason why ECC technology currently in wide use in server memory, it is because in the past other new memory technology is not yet mature, in addition to the current server in the system speed is still very high, at this frequency generally occur at multiple bit errors phenomenon rarely occurs, because this allows ECC technology is fully recognized and applied, making the ECC memory technology into almost all the memory on the server.

But with the Intel processor-based architecture of the server's CPU performance in a multiple of the geometry class, and the hard drive's performance during the same period only improves a few times, so in order to obtain adequate performance, the server requires a large amount of memory to temporarily save the CPU needs to read data, such a large amount of data access results in a single memory chip, on every visit is usually to provide 4 (32-bit) or 8 (64-bit) bit more data to read so much data at once, the possibility of many data errors will greatly improve and ECC cannot correct double bit error above, this is likely to cause all the bits of data loss, the system will soon collapse. IBM Chipkill technology is the use of the memory of the substructure method to address this challenge. Memory subsystem design principle is that a single chip, no matter how much data width is only for a given ECC identification number, its impact up to a bit. For example to illustrate is that if you use a 4-bit wide DRAM, 4 bits each bit parity will be composed of different identification code, the ECC ECC identification number is used to separate a bit to save the data, that is saved in a different memory address space. Therefore, even if the entire memory chips down each ECC identification number will also be up to a bit of bad data, but this situation can be fixed by ECC logic, thus ensuring fault tolerance of the memory subsystem to ensure that the server in the event of a failure, there is a strong resiliency. This memory technology memory can also check and fix data bit 4 errors, server reliability and stability have been more fully protected.

(3) Register

Register is a register or registers, in-memory directory on the understanding that we can put it into the contents of the book, with it, when memory read and write instructions received, will first search this directory, and then read and write operations, which will greatly improve server memory efficiency. With the memory must Register with Buffer (buffer), and currently can see Register memory also have ECC function, it is mainly used in the high-end server and graphics workstation, such as IBMNetfinity5000.

(4) FB-DIMM

FB-DIMM (FullyBuffered-DIMM, fully buffered memory module) is Intel in DDR2, DDR3, based on the development of a new type of memory module and the Internet architecture, you can mix now DDR2 memory chip that can match the future DDR3 memory chips. FB-DIMM can greatly enhance the system memory bandwidth and greatly increase memory maximum capacity.

FB-DIMM technology is Intel to address memory performance on system overall performance of the systemAbout developed, on the basis of existing technologies to achieve a rapid increase in performance, while costs are relatively low. In the entire computer system, the memory can determine the key factors of machine performance, a faster CPU, no good memory systems and complexes, CPU performance and excellent also cannot play. This situation is determined by the principles of computer, CPU in operation when required data is obtained from the memory, if the memory system is unable to supply data to the CPU, the CPU must longer time in a wait state, hardware resource idle, performance will not play. For an ordinary personal computer, as a single processor system, the current memory bandwidth has been able to meet their performance requirements; for multiple servers, because it is a multiprocessor system, its memory bandwidth and memory capacity is extreme craving, traditional memory technology has been unable to meet their needs. This is because the current normal DIMM is a kind of "short-term connection" (Stub-bus) topology, this kind of structure, each chip and memory controller data bus has a short circuit, causing a impedance not continue, thus affecting the stability and integrity of the signal, the higher the frequency the more data or chip, the greater the impact. Although the company launched the Rambus XDR memory and other new memory technology with extremely high performance, but there is a question of cost is too high, so that they are not universal. But there are techniques of FB-DIMM better solved the problem, not only provides greater memory capacity and better memory bandwidth and remained relatively low cost. Compare with XDR FB-DIMM, although the performance is not as new XDR schema, but the cost is low compared to XDR.

And existing ordinary DDR2 memory technology has compared FB-DIMM great advantages: in memory of the same frequency currently offers four times the normal memory bandwidth, and can support a maximum memory capacity reached ordinary memory of 24 times, maximum system memory can support 192GB. FB-DIMM is characterized by the largest adopts existing DDR2 memory chip (later also uses DDR3 memory chips), but with a PCB memory buffer chip AMB (Advanced memory buffer AdvancedMemoryBuffer) converts the parallel data through the serial data streams like PCIExpress point-to-point high-speed serial bus transfers data to the processor.

With normal DIMM module technology FB-DIMM and data between the memory controller and command transmission is no longer the traditional design of parallel lines, and has adopted a similar to PCI-Express serial interface design of multi-channel parallel to serial manner the data transfer. In this new framework, each DIMM is buffer on each line, is a point-to-point connection between the way that the data will be passed after the first buffer passed to the next buffer, such that the first buffer and memory controllers connect impedance can remain stable, thus contributing to the improvement of the capacity and frequency.

Current server commonly used type of memory is DDR2 SDRAM and DDR, three memory.

Because the server memory in various technologies relative compatible, will be much more stringent, it emphasizes not only the memory speed, but its internal error correction technology capacity and stability. So far on the FSB is immediately compatible or normal desktop memory. Current desktop FSB in General has reached more than 150MHz times, but outside the mainstream 133 is frequent. While the server because the entire accessories FSB and high stability requirements, mainstream FSB or 100MHz, but 133MHz FSB has gradually in all grades server implementation, when buying, of course, it is best to choose the server 133MHz FSB! memory, other accessories too, to try to sync to, otherwise it will affect the entire server's performance. At present, the main server memory brand mainly Kingmax, kinghorse, Hyundai, Samsung, IBM, kingstone, VIKING, NEC, etc, but mainly in front of some of the more common in the market, and the quality will be better protected.