8b_10b_编码

8b/10b编码 2007-04-04 22:13 8b/10b编码是目前许多高速串行总线采用的编码机制，如1394b、Serial ATA、PCI Express、Infini-band、Fiber Channel、RapidIO等总线或网络，都是8b/10b编码的拥护者。8b/10b编码方式早在1983年就被提出，现在则成为主流传输选用 的编码技术，本文将探究其优点，并详细介绍8b/10b编码重要的三大观念。 以映射机制将8位转为10位字码 　　8b/10b编码是目前高速数据传输接口或总线常用的编码方式，该编码技术的基本精神很简单，就是将8个位经过映射的机制转化为10个位的字 码，但是平衡了位流中0与1的数量，这就是8b/10b编码的根本目的“直流平衡（DC Balance）”。当高速串行流的逻辑1或逻辑0有多个位没有产生变化时，信号的转换就会因为电压位阶的关系而造成信号错误，直流平衡的最大好处便是可 以克服以上问题。 　　80年代IBM发表该论文时，正是预期到光纤技术的应用实现而针对高速局域网以及计算机链接所研发的传输码。事隔多年，8b/10b 编码竟摇身一变成为高速串行总线的主要编码方式，这也许是当时所始料未及的。 　　8b/10b编码方式存在着一个极为有趣的特色，其总输出位数是10个位，其中位0与位1出现的次数总共也仅在三种场合存在，分别为“5个位0与5个位1”、“4个位0与6个位1”、“6个位0与4个位1”。 　　有一个8b/10b编码的技术专用语“不均等性（Disparity）”，其涵义就是指10个位中位0与位1出现次数的差。换句话说， 8b/10b编码的“Disparity”就仅有“+2”（4个位0与6个位1）、“+0”（5个位0与5个位1）以及“-2”（6个位0与4个位1）三 种状况。 分两个部分进行映射处理 　　从图1至少可以看出几点关于8b/10b编码的基本观念。首先，8个位映射成10个位，划分两个部分分别进行映射处理，分别是5B/6B与 3B/4B的处理。这种位的映射关系，已经在当时论文中做成一个标准化的。其次，5B/6B与3B/4B的映射控制，必须依赖“Disparity Control”功能块才能够达成最后的编码，因为直流平衡就是要靠它制衡。 　　另一个重点是，当初在规划8b/10b编码时，除了“数据码”的定义之外，又另外规划了所谓的“控制代码”。文章或是文献上所见到的 “D.a.b”就是数据码的表现方式。而“K.a.b”就是另指控制代码，至于为何会有“D.a.b”及“K.a.b”这种表现方式，当然与前面所提到的 5B/6B与3B/4B的映射控制有高度相关，也有人以“/Dx.y/”或是“/Kx.y/”的文字来分别表示。 　　一个字节的组成有8个位，若是将其分成两类，分别是5个LSB以及3个MSB。5个LSB就是前面提及的“x”，3个MSB就是“y”。以下举例说明这种表现方式。 　　示范之一，/D0.0/ = 000 00000 　　示范之二，/D6.2/ = 010 00110 　　如此一来，3B/4B、5B/6B的各自映射就更容易呈现。 　　由于高速串行流是连续地将数据从发送端往接收端逐次地传输过去，每一个数据的编码会随着前面送出数据不停累积“不均等性”的状态而产生正确的位 映射关系，所以这个依靠做为判断的状态就被称为“Running Disparity”。实际上可以将它视为一个判断标志（Flag），因此“Running Disparity”也仅会出现两种状态，分别是“+1”与“-1”，分别代表位1比位0多或位0比位1多。如此一来，整个冗长数据串的编码就可以被限制 在约定的范围之内。 数据码与控制代码不会同时发生 　　表2、表3列举了5B/6B、3B/4B的编码映像表格。当初在规划8B/10B编码机制的时候，就事先规划了所谓的控制代码（Control Characters），也有人称为“K Code”，顾名思义，就是作为控制的用途。由于数据码与控制代码两者不会在同一个时间点发生，因此即便数据码与控制代码在编码上的数值重复，也不会发生 冲突。 　　控制代码的另一个重点是：这些控制代码会随着各种接口的需要而有各自不同的解读与定义，以下就以流行的Serial ATA与PCI Express为例来解释这个要点。 　　先来看PCI Express如何来运用这些控制代码（如表5所示）。PCI Express运用控制代码基本上作为两种用途：其一，接口总线上的链接管理（Link Management）必须依靠它；其二，链接层与数据交易层各自生成的包（DLLP与TLP），发送的开始与退出也需要控制代码来标示。而致力于存储装 置接口的Serial ATA规范，则是参考Fibre Channel的内在，仅仅用到K28.3与K28.5。 　　如果从元器件的角度来看 8b/10b的功能，究竟藏身何处？答案也是很清楚，就在物理层（Physical layer）当中。请看PCI Express的PHY逻辑示意图（图5），自然能够理解串行总线的一些基本法则。 　　总之，8b/10b编码机制的发源虽然有点久远，但却对高速串行总线具有深远的影响，并成为现今开发人员的重要课题。 Data Encoding Techniques Introduction In order to transport digital bits of data across carrier waves, encoding techniques have been developed each with their own pros and cons. This document briefly describes some of the more common techniques. Character Coding Techniques Baudot Jean-Maurice-Emile Baudot developed a character set in 1874 that used series of bits to represent characters that could be sent over a telegraph wire or radio signal. A 5-key keyboard was developed to implement this Baudot code that was modified by Donald Murray in 1901 and it became the International Telegraph Alphabet 1 (ITA1) and then developed into ITA2. ITA2 was the coding that was actually implemented on equipment. Characters such as Line Feed (LF) were given a 5 bit code such as 00010. The problem with using 5 bits for each character is that there is a limitation on the number of characters that can be generated from them, 25 gives 32 different combinations. This may be fine for 26 letters of the English alphabet but it is not enough to cover punctuation or control characters. Other coding techniques were needed. Binary Coded Decimal (BCD) BCD uses a series of 4 bits called a nibble to represent a decimal number, as the following table demonstrates: Decimal 0 1 2 3 4 5 6 7 8 9 BCD 0000 0001 0010 0011 0100 0101 0110 0111 1000 1001 So for example, the number 1456 would be represented by 0001 0100 0101 0110. This makes it easier to convert numbers and for displays, however the electronics required in calculations is quite complex. American Standard Code for Information Interchange (ASCII) Originally published in 1963, ASCII is based on 7 bits to represent English characters and after a number of revisions ASCII now supports 95 printable characters and 33 control characters (a total of 2 7 = 128). ASCII is the americanised vsersion of that defined by CCITT in ISO 646 and is known as the International Alphabet 5 (IA5). The first 32 characters are control characters and are represented by the 7-bit codes 000 0000 (null character) through to 001 1111 (unit separator). The 128th control character is 'delete' represented by 111 1111. The rest of the characters are printable and the coding caters for both lower and uppercase english letters e.g. the letter 'd' is represented by 110 0100 whereas its upper case equivalent is represented by 100 0100. Extended Binary Coded Decimal Interchange Code (EBCDIC) Around the same time that ASCII was developed, in 1964 IBM produced EBCDIC which is an 8-bit coding system designed to replace BCD within its computer systems. An EBCDIC byte is divided in two nibbles. The first four bits is called the zone and this represents the category of the character, the last four bits is called the digit and this identifies the specific character. Different countries adapted EBCDIC for their own alphabets. The Chinese had a double byte extension that allowed them to display Chinese characters. IBM numbered the different character sets with Coded Character Set Identifier (CCSID) of which there are many around the world. Unicode Originally published in 1991 by the Unicode Consortium as Unicode 1.0 (in 2006 Unicode 5.0 was released), Unicode aims to provide a means for the traditional character sets around the world to take part in multilingual computer processing amongst themselves rather than have to translate into a Roman character set first. The bit patterns of the 95 printable ASCII characters are sufficient to exchange information in modern English, however many languages that use the Latin alphabet need additional symbols not covered by ASCII. ISO/IEC 8859 attempts to address this by utilising the eighth bit in an 8-bit byte in order to allow positions for another 128 characters. This bit was previously used for data transmission protocol information, or was left unused. Even more characters were needed than could fit in a single 8-bit character encoding, so several mappings were developed. ISO/IEC 8859 comes in parts and these are given a number e.g. ISO 8859-15. Unicode creates codes for the characters or basic graphical representation of the character (called a 'grapheme'). The first 256 code points have been reserved for ISO 5589-1 in order to make it straightforward to convert the Roman text. There are two Unicode mapping methods; Unicode Transformation Format (UTF) and Unicode Character Set (UCS). An encoding maps the range of Unicode code points to sequences of values in a fixed-size range of code values. The numbers in the names of the encodings indicate the number of bits in one code value (for UTF encodings) or the number of bytes per code value (for UCS) encodings. UCS assigns a code per character. UCS-2 uses two bytes per character, UCS-4 uses 4 bytes per character. Some Unicode examples: · UTF-7 — a 7-bit encoding, often considered obsolete (not part of Unicode but rather an RFC) · UTF-8 — an 8-bit, variable-width encoding, which maximizes compatibility with ASCII. In common use and is in fact a superset of ASCII. The IMC and IETF use UTF-8 when determining standards for supporting email and Internet traffic. · UTF-EBCDIC — an 8-bit variable-width encoding, which maximizes compatibility with EBCDIC. (not part of Unicode) · UTF-16 — a 16-bit, variable-width encoding. In common use. · UTF-32 — a 32-bit, fixed-width encoding Manchester Phase Encoding (MPE) 802.3 Ethernet uses Manchester Phase Encoding (MPE). A data bit '1' from the level-encoded signal (i.e. that from the digital circuitry in the host machine sending data) is represented by a full cycle of the inverted signal from the master clock which matches with the '0' to '1' rise of the phase-encoded signal (linked to the phase of the carrier signal which goes out on the wire). i.e. -V in the first half of the signal and +V in the second half. The data bit '0' from the level-encoded signal is represented by a full normal cycle of the master clock which gives the '1' to '0' fall of the phase-encoded signal. i.e. +V in the first half of the signal and -V in the second half. The above diagram shows graphically how MPE operates. The example at the bottom of the diagram indicates how the digital bit stream 10110 is encoded. A transition in the middle of each bit makes it possible to synchronize the sender and receiver. At any instant the ether can be in one of three states: transmitting a 0 bit (-0.85v), transmitting a 1 bit (0.85v) or idle (0 volts). Having a normal clock signal as well as an inverted clock signal leads to regular transitions which means that synchronisation of clocks is easily achieved even if there are a series of '0's or '1's. This results in highly reliable data transmission. The master clock speed for Manchester encoding always matches the data speed and this determines the carrier signal frequency, so for 10Mbps Ethernet the carrier is 10MHz. Differential Manchester Encoding (DME) A '1' bit is indicated by making the first half of the signal, equal to the last half of the previous bit's signal i.e. no transition at the start of the bit-time. A '0' bit is indicated by making the first half of the signal opposite to the last half of the previous bit's signal i.e. a zero bit is indicated by a transition at the beginning of the bit-time. In the middle of the bit-time there is always a transition, whether from high to low, or low to high. Each bit transmitted means a voltage change always occurs in the middle of the bit-time to ensure clock synchronisation. Token Ring uses DME and this is why a preamble is not required in Token Ring, compared to Ethernet which uses Manchester encoding. Non Return to Zero (NRZ) NRZ encoding uses 0 volts for a data bit of '0' and a +V volts for a data bit of '1'. The problem with this is that it is difficult to distinguish a series of '1's or '0's due to clock synchronisation issues. Also, the average DC voltage is 1/2V so there is high power output. In addition, the bandwidth is large i.e. from 0Hz to half the data rate because for every full signal wave, two bits of data can be transmitted (remember that with MPE the data rate equals the bit rate which is even more inefficient!) i.e. two bits of information are transmitted for every cycle (or hertz). After 50m of cable attenuation the signal amplitude may have been reduced to 100mV giving an induced noise tolerance of 100mV. Return to Zero (RZ) With RZ a '0' bit is represented by 0 volts whereas a '1' data bit is represented by +V volts for half the cycle and 0 volts for the second half of the cycle. This means that the average DC voltage is reduced to 1/4V plus there is the added benefit of there always being a voltage change even if there are a series of '1's. Unfortunately, the efficiency of bandwidth usage decreases if there are a series of '1's since now a '1' uses a whole cycle. Non Return to Zero Invertive (NRZ-I) With NRZ-I a '1' bit is represented by 0 volts or +V volts depending on the previous level. If the previous voltage was 0 volts then the '1' bit will be represented by +V volts, however if the previous voltage was +V volts then the '1' bit will be represented by 0 volts. A '0' bit is represented by whatever voltage level was used previously. This means that only a '1' bit can 'invert' the voltage, a '0' bit has no effect on the voltage, it remains the same as the previous bit whatever that voltage was. This can be demonstrated in the following examples for the binary patterns 10110 and 11111: Note how that a '1' inverts the voltage whilst a '0' leaves it where it is. This means that the encoding is different for the same binary pattern depending on the voltage starting point. The bandwidth usage is minimised with NRZ-I, plus there are frequent voltage changes required for clock synchronisation. With fibre there are no issues with power output so a higher clock frequency is fine whereas with copper NRZ-I would not be acceptable. 4B/5B 4B/5B encoding is sometimes called 'Block coding'. To get around this problem, an intermediate encoding takes place before the MLT-3 encoding. Each 4-bit 'nibble' of received data has an extra 5th bit added. If input data is dealt with in 4-bit nibbles there are 24 = 16 different bit patterns. With 5-bit 'packets' there are 25 = 32 different bit patterns. As a result, the 5-bit patterns can always have two '1's in them even if the data is all '0's a translation occurs to another of the bit patterns. This enables clock synchronisations required for reliable data transfer. Notice that the clock frequency is 125MHz. The reason for this is due to the 4B/5B encoding. A 100MHz signal would not have been enough to give us 100Mbps, we need a 125MHz clock. 5B/6B Same idea as 4B/5B but you can have DC balance (3 zero bits and 3 one bits in each group of 6) to prevent polarisation. 5B/6B Encoding is the process of encoding the scrambled 5-bit data patterns into predetermined 6-bit symbols. This creates a balanced data pattern, containing equal numbers of 0's and 1's, to provide guaranteed clock transitions synchronization for receiver circuitry, as well as an even power value on the line. 5B6B encoding also provides an added error-checking capability. Invalid symbols and invalid data patterns, such as more than three 0's or three 1's in a row, are easily detected For 100VG-AnyLAN for instance, the clock rate on each wire is 30MHz, therefore 30Mbits per second are transmitted on each pair giving a total data rate of 120Mbits/sec. Since each 6-bits of data on the line represents 5 bits of real data due to the 5B/6B encoding, the rate of real data being transmitted is 25Mbits/sec on each pair, giving a total rate of real data of 100Mbits/sec. For 2-pair STP and fiber, the data rate is 120Mbits/sec on the transmitting pair, for a real data transmission rate of 100Mbits/sec. 8B/6T 8B/6T means send 8 data bits as six ternary (one of three voltage levels) signals. 3/4 (6/8) wave transitions transitions per bit i.e. the carrier just needs to be running at 3/4 of the speed of the data rate. The incoming data stream is split into 8-bit patterns. Each 8-bit data pattern with two voltage levels 0 volts and V volts is examined. This 8-bit pattern is then converted into a 6-bit pattern but using three voltage levels -V, 0 and V volts, so each 8-bit pattern has a unique 6T code. For example the bit pattern 0000 0000 (0x00) uses the code +-00+- and 0000 1110 (0x)E) uses the code -+0-0+. There are 36 = 729 possible patterns (symbols). The rules for the symbols are that there must be at least two voltage transitions (to maintain clock synchronisation) and the average DC voltage must be zero (this is called 'DC balance' that is the overall DC voltage is summed up to 0v, the +V and -V transitions are evenly balanced either side of 0V) which stops any polarisation on the cable. The maximum frequency that the 6T codes could generate on one carrier is 37.5MHz. FCC rules do not allow anything above 30MHz on cables and Category 3 cable does not allow anything above 16MHz (which is what 100BaseT4 was designed for). The 100BaseT4 standard uses 8B/6T encoding on three pairs in a round robin fashion such that the maximum carrier frequency on any single pair is 37.5/3 = 12.5MHz. 8B/10B Each octet of data is examined and assigned a 10 bit code group. The data octet is split up into the 3 most significant bits and the 5 least significant bits. This is then represented as two decimal numbers with the least significant bits first e.g. for the octet 101 00110 we get the decimal 6.5. 10 bits are used to create this code group and the naming convention follows the format /D6.5/. There are also 12 special code groups which follow the naming convention /Kx.y/. The 10 bit code groups must either contain five ones and five zeros, or four ones and six zeros, or six ones and four zeros. This ensures that not too many consecutive ones and zeros occurs between code groups thereby maintaining clock synchronisation. Two 'commas' are used to aid in bit synchronisation, these 'commas' are the 7 bit patterns 0011111 (+comma)and 1100000 (-comma). In order to maintain a DC balance, a calculation called the Running Disparity calculation is used to try to keep the number of '0's transmitted the same as the number of '1's transmitted. This uses 10 bits for each 8 bits of data and therefore drops the data rate speed relative to the line speed, for instance in order to gain a data rate of 1Gbps the line peed has to be 10/8 x 1 = 1.25Gbps . MLT-3 This scheme was specified by ANSI X3T9.5 committee. It is used by FDDI and TP-PMD to obtain 100MB/s out of a 31.25MHz signal. UTP is low pass in nature, meaning that it hinders high frequency signal (like a low-pass filter). So it is not feasible to merely increase the clock frequency by 10 to 100MHz and use Manchester encoding to give us 100Mbps. In addition, the FCC (Federal Communications Commission) have severely curtailed the power that is allowed to be emitted above 30MHz. We have to use another encoding technique in order to transmit high data rates across UTP. If you take an averaging spectrum analyser and look at the output signal of the 10Mbps Ethernet phase-encoded signal, you will see a power peak at 10MHz where there is a stream of '1's or '0's, you will see a smaller harmonic at 30MHz and if there is a stream of '1's and '0's, you will see a peak at 5MHz. Now 100BaseT uses a master clock running at 125MHz instead of 10MHz. The equivalent peaks would then be at 125MHz, 375MHz and 62.5MHz. Transmission electronics designed to work within the FCC rules will block the frequencies higher than 30MHz. To get around this issue we need to concentrate the signal power below 30MHz if possible. To do this the encoding method Multi-Level Transition 3 (MLT-3) is used. This involves using the pattern 1, 0, -1, 0. If the next data signal is a '1' then the output 'transitions' to the next bit in the pattern e.g. if the last output bit was a '-1', and the input bit is a '1', then the next output bit is a '0'. If the next data signal is a '0' then there is no transition which means that the next output bit is the same as last time, in our case a '0'. The cycle length of the output signal is therefore going to be 1/4 that of the MPE method so that instead of the main signal peak being at 125MHz as measured by the averaging spectrum analyser, it will be at 31.25MHz which is near enough to be OK as far as FCC are concerned. 5 bits are transmitted for every 4 bits of data so that the data bit rate is actually 125Mb/s for 100Mb/s data throughput. There is an issue with this in that you can end up with a series of '0's or '1's which force the local circuitry to count the bits using its own free running clock rather than have the check of the clock synchronisation from the transmit source. PAM-5 This employs multi-level amplitude signalling. To encode 8 bits, 28 = 256 codes or symbols, are required since there are 256 possible pattern combinations. A five level signal (e.g. -2v, -1v, 0v, 1v and 2v) called Pulse Amplitude Modulation 5 is used (This works in a similar manner to MLT-3). Bearing in mind that there are 4 separate pairs being used for transmission and reception of data, this gives us a possibility of 54 = 625 codes to choose from when using all four pairs. Actually only four levels are used for data, the fifth level (0v) is used for the 4-dimensional 8-state Trellis Forward Error Correction used to recover the transmitted signal from the high noise. If you plot time (nanoseconds) against voltage you will see an 'eye pattern' effect showing the different signal levels. Comparing a plot for MLT-3 against PAM-5 will demonstra