Many impressive feats are possible with high-speed networking -- most involving transmission of high-quality video and audio services -- but many of these require sophisticated cabling arrangements that are found chiefly in commercial networks. What are the options for bringing these services withinthe walls of our homes?
A major obstacle to such deployments is that the phone lines found in most residences are of inconsistent quality for high-speed, high-quality networking. The Home Phoneline Networking Alliance (HomePNA) is addressing the problem with networking specifications designed to deliver high-speed networking capabilities to residences over impaired channels.
The first release of the standard operates at a speed of 1 Mbps. The initial release of the second HomePNA standard, version 2.0, was initially designed for a throughput of 16 Mbps (marketed at 10 Mbps) but is now poised to deliver 32-Mbps transmissions. Only slight changes to existing chipsets are necessary because initial HomePNA 2.0 products were designed from inception to scale to higher speeds using improved modulation and transceiver
techniques.
For this transition to take place, consumers must be able to connect their home gadgets with a shared digital communication line. The idea of stringing coaxial cable throughout homes is too disruptive as well as too expensive to be practical. Therefore, three potential infrastructures have emerged: phone-line wiring, wireless, and AC power wiring. While it appears that all three will eventually be used together, phone-line networks are being deployed first and have sold more than 2 million chipsets to date -- by far, the most widely adopted of the three methods.
In 1998 the computer and semiconductor industries created HomePNA to select, promote, and standardize technologies for home phone-line networking. In
developing the specification, the group had to take into account that networking over the existing home phone lines suffers from many impairments (as do
all no-new-wires physical media), namely high attenuation, reflections, impulse noise, crosstalk, and RF ingress from nearby transmitters.
Moving forward
The first version of the HomePNA's standard was designed to support 1-Mbps transmission over these lines (see Figure 1). HomePNA 2.0 was built to work first with 16-Mbps transmissions, and is now doubling that speed to 32 Mbps. By examining the standard's frequency range and modulation technique, one can understand how the performance boost occurs.
One of the first decisions the alliance faced was selection of the proper frequency range. The 4- to 10-MHz band was chosen for several reasons. The lower limit of 4 MHz makes it feasible to implement the filters needed to reduce out-of-band interference between HomePNA devices and other products such as asymmetric digital subscriber line (ADSL) modems. After modeling several thousand representative networks with telephones and common wire lengths, it was determined that the spectrum above 10 MHz was much more likely to have wider and deeper nulls caused by reflections.
Cutting down crosstalk
Crosstalk between phone lines was also an important concern. This increases with frequency, and the analog front end (AFE) is harder to implement at higher frequencies. In comparison, the 4- to 10-MHz range only overlaps a single amateur radio band (40 m), which simplifies ingress and egress filtering.
The next consideration was a modulation technique. HomePNA 1.0 uses a pulse position modulation (PPM) technique that supports a 1-Mbps data rate. The second-generation system targets an order-of-magnitude higher data-rate improvement. The use of quadrature amplitude modulation (QAM) in the HomePNA 2.0 specification provides the ability to get more throughput in the same bandwidth and to achieve greater robustness.
Because home phone-line channels often have very deep nulls, and multiple nulls in band, a few additions were needed. The first was adapting the modulation rate. Instead of having a fixed number of bits per symbol, a transmitter may, on a packet-by-packet basis, vary the packet encoding from 2 to 8 b per symbol. A packet header is always encoded at 2 b per symbol, so that every receiver can demodulate at least the packet's header.
The HomePNA standard uses a fixed 7-MHz carrier frequency and can operate at either 2 or 4 Mbaud with modulation encodings of 2 to 8 b per symbol. The base symbol rate is 2 Mbaud. At this rate, the standard has a peak data rate ranging from 4 to 16 Mbps, though overhead reduces the actual throughput the HomePNA standard can achieve.
Unfortunately, the nature of channel nulls can be such that even rate adaptation down to 2 b per symbol is not sufficient to guarantee that the packet can be received. In a traditional QAM system, if there is an extreme null (that is, one with which the equalizer cannot cope) in the band, the system will fail to operate.
So, at its 2-Mbaud rate, Home-PNA 2.0 implements a modified version of QAM known as frequency-diverse QAM (FD-QAM).
In a traditional QAM system, a single copy of the baseband signal is sent and received. Because in FD-QAM the symbol rate is less than half the filter's width, the output signal has two redundant copies of the baseband signal. Thus, the signal is frequency diverse, leading to the name FD-QAM.
Intuitively, it is easy to see that on channels where half of the spectrum is nulled out, one copy of the signal will still make it through. This design was followed due to line-quality concerns. However, results from HomePNA 2.0 deployments have demonstrated that in many installations, the second data copies are seldom necessary. Now, the transmission speed can be doubled by simply replacing the redundant data with more information, which is the course being followed to deliver 32-Mbps transmissions.
Yet, the specification is still able to field quite a bit of distortion. While there is no forward error correcting (FEC) mechanism, it does include fast automatic repeat request features.
Two key elements
Higher speeds require vendors to design receivers with precision so that noise distortions will not inhibit performance. Design verification for this precision involves two tools, the first of which is a channel-modeling system that was developed based on measuring actual homes, combined with theoretical models for parameters such as wire gauge and insulation materials. This channel model allows us to generate "cooked" packets, which lets a tester create a set of digital samples corresponding to a packet that traversed a particular channel from one node to another.
The other key element is a software environment for rapidly developing an implementation of the transmitter and receiver. In its simplest mode, the verification system could be used to transmit a packet, which is then cooked using the channel model and demodulated using the receiver. In practice, the transmitter implementation is much simpler than the receiver. Using the channel modeler, the simulated chipset could generate millions of packets (thousands of packets cooked by thousands of network topologies) to verify that the receiver was operating correctly. In addition, testers hand-generated several worst-case packets that had, for example, worst-case peak-to-average signals, timing offset, and so on.
The next important use of the software implementation of the system is the creation of a soft-physical layer (PHY). In this mode, together with a hardware AFE board that plugs into a PC, we were able to send and receive HomePNA 2.0 signals at a reasonable rate, prior to silicon availability. By using Pentium MMX instructions to augment certain inner loops, the specification demodulates at rates greater than 4 Mbps on a single-processor 400-MHz Pentium.
This soft implementation supports verification of the system's correct operation, debugging the upper layers of the drivers and the link-layer protocols such as limited automatic repeat request (LARQ), and perhaps most important, providing realistic and impressive demonstrations of the system. The software implementation was also instrumental in debugging the register transfer logic (RTL) used to implement the HomePNA 2.0 chipset (a generic term that could be used by any manufacturer; Broadcom happens to be the first one).
Connecting simulation
Finally, the verification environment provided the means to connect multiple instances of either hardware or software simulations verifying a correctly functioning media access control (MAC). This is important because the primary difference between twisted-pair Ethernet and other technologies is the quality of the communication channel. Running over category-5 cable, Ethernet encounters a channel that has a number of favorable properties. They include point-to-point communication, proper termination, a well-characterized channel response (both in terms of nulls and overall attenuation), and
very low crosstalk. In contrast, all of the no-new-wires media available for networking within homes often have a severely impaired communication
channel.
A fundamental design challenge for home phone-line networking is how a receiver can determine - for each burst transmission seen on the channel - the equalization and demodulation parameters that should be applied to recover the packet. And conversely, for each transmitter, the modulation rate that is feasible when sending to a particular destination station must be determined. In HomePNA 2.0, this is accomplished with a self-describing frame format with PHY-level signals that can directly control equalizer training and demodulation as well as a rate negotiation algorithm.
Sticking to standards
The importance of leveraging standards cannot be overestimated. There are two issues to consider. First, is there an accepted standard that guarantees interoperability between equipment from multiple manufacturers, and second, does the network faultlessly support other networking standards and in particular the IP suite. Given the preponderance of IEEE-802.3 Layer 2 networking across the Internet infrastructure, HomePNA chose technology that uses 802.3 framing and Ethernet carrier-sense multiple access with collision detection (CSMA/CD) MAC behavior.
HomePNA 2.0 describes a multipoint CSMA/CD packet network that supports unicast, multicast, and broadcast transmissions.
It has the look and feel of Ethernet. However, it differs from 10Base-2 and 10Base-T in a number of re-spects. First and foremost, Home-PNA 2.0 does not place any restrictions on wiring type, wiring topology, or termination. Moreover, like 10Base-2, but unlike 10Base-T, HomePNA 2.0 uses a shared physical medium with no need for a switch or hub. On the other hand, 10Base-T requires dedicated point-to-point Category-3 or Category-5 wires. Total cost of ownership for the average consumer is far lower than conventional Ethernet that requires new wiring, complex hubs and technical configuration.
At the PHY, the system is frequency-division multiplexed (FDM) on the same wire as used by the standard analog phone service and splitterless ADSL. Analog telephony uses the low part of the spectrum, below 26 kHz. ADSL (both G.Lite and G.Heavy) use spectrum up to 1.1 MHz.
HomePNA frame format
The basic HomePNA 2.0 frame format shows the frame format on the wire (see Figure 2). Each frame begins with a known 64-symbol preamble. The preamble supports robust carrier sensing and collision detection, equalizer training, timing recovery, and gain adjustment. Following the preamble is a frame control field, the first part of which is an 8-b frame type. Frame-type = 0 is shown in the figure, where other codes can be assigned for future system frame formats.
Following the frame type is an 8-b field that specifies the modulation format (bits per symbol, for example). There are other miscellaneous control fields in frame control including an 8-b cyclic redundancy check (CRC) header. The remainder of the packet is exactly an 802.3 Ethernet frame followed by CRC16, padding, and end-of-file (EOF) sequence. The CRC16 covers the header and payload, and reduces the undetected error rate for severely impaired
networks.
Key to operation is that the first 120 b of the frame are sent at the most robust 2-Mbaud, 2-b/symbol rate so that any station able to demodulate a packet can do it at this encoding. Thus, even if the payload is encoded at a rate that the receiver cannot demodulate, it will be possible to demodulate the header. In this situation, the receiver sends a rate-request control frame (RRCF) to the sender, asking it to reduce the number of bits per symbol or the symbol
rate.
In practice, the system starts out sending at 2 b/symbol and 2 Mbaud unless the receiver sends an RRCF, asking for future packets to be sent at higher data rates. Several algorithms can be used to determine when to send RRCFs and to estimate the channel capacity using an approximate signal-to-noise ratio (SNR) and bit-error statistics. The rate-adaptation algorithm can optimize the rate used when sending to multicast and broadcast groups.
QoS and latency planning
Anticipated home applications will drive the requirement for quality of service (QoS) support. The first applications of home networking will involve relatively simple PC-only functions, such as Internet access, file sharing, and printer servers. In a few years, applications based on digital audio, digital video, and digital voice (IP telephony) will gain acceptance.
Latency in voice connections must be controlled on the home network segment if voice quality is to be maintained. Streaming video and audio connections must be guaranteed minimum amounts of bandwidth in order to deliver high-quality transmissions. Without a QoS mechanism, burst loads generated by various applications would make the network unable to meet the latency and guaranteed bandwidth service requirements at times.
HomePNA 2.0 introduces eight levels of priority and uses a new collision-resolution algorithm known as distributed fair priority queuing (DFPQ). Voice telephony requires a low-latency network service, and streaming audio or video applications require a guaranteed bandwidth service. With the MAC in Ethernet, there are no real service guarantees.
For example, three nodes (N0, N1, and N2) could contend for access to the network. Node N2 is transmitting a voice-over-IP (VoIP) packet. Initially, N0 accesses the wire and transmits a frame (TX). During this transmission N2 has packetized a voice sample and is ready to transmit, but must defer to N0. At the end of the first transmission, N0 has a second packet ready to send, and when N2 and N0 contend for access (resulting in a collision), N2 by chance chooses a longer back-off interval than N0. N0 gains access again and transmits. During this time, another station N1 becomes active and starts deferring,
waiting for N0 to finish. Now, when N2 attempts to transmit, it collides with N1. A possible outcome is that N1 succeeds in the collision resolution, and N2 further increases its back-off.
When a binary exponential back-off is used with counter reset on successful transmission, the queuing discipline can become very unfair for N2. If N0 and N1 are PCs engaged in file transfers, they can generate enough traffic loading on the network to cause dropped frames in the VoIP service operating on N2.
One solution is to introduce different access priorities where the VoIP station uses a higher priority than best-effort file-transfer traffic. HomePNA 2.0 accomplishes this by organizing the time following the interframe gap into an ordered series of priority slots (see Figure 3). When N0 finishes
transmitting, all stations on the network with a lower priority than 7 wait while N2 begins to transmit (without collision). After N2's transmission, no stations have traffic with priority higher than 1, so N0 again gains access to the channel with its next transmission.
Access priority lets software define different service classes such as low latency, controlled bandwidth, guaranteed bandwidth, best effort, and penalty. Each uses a different priority level. Within a given priority level, HomePNA 2.0 uses a new algorithm for collision resolution.
Each station keeps track of a back-off level and after a collision, randomly chooses to increment the back-off level by 0, 1, or 2. During a collision-resolution cycle, stations incrementally establish a partial ordering. Soon, only one station remains at the lowest back-off level and gains access to the channel.
In the Figure 4 example, under Conflict Resolution HomePNA collision resolution, N0 and N1 enter into a collision resolution cycle. N0 randomly chooses to increment its back-off level by 2, and N1 by 0. To optimize the partial ordering, eliminating null levels, stations send a special signal immediately following a collision. The signal reflects the back-off increment chosen (0 and 2 in the example shown).
All stations observe these signals and perform a distributed computation to calculate the new (partial) ordering. In this case, N0 increments its back-off level by 1 because it saw the back-off signal from N1 in S0, but no station indicating in S1.
Channel impairments
Impulse noise is another impairment for all home networks using no-new-wires technology. On phone wires, this noise is often due to phone ringing, switch-hook transitions, and noise coupled from the AC power wiring. Fortunately, the impulses tend to be short and destroy only a single packet. While there are coding techniques that might reduce the number of packets destroyed by impulses, HomePNA has chosen to use a fast retransmission mechanism designated LARQ. LARQ is preferable to fast Ethernet connections for relatively low frame-erasure rates.
Since LARQ is implemented (in software) at Layer 2 and because it operates only on a single segment of the network, it is effective in hiding frame erasure from TCP/IP (see Figure 5).
Finally, note that HomePNA 2.0 implements a link-integrity mechanism, which can be implemented either in hardware or at low levels of a software driver. The virtue of link integrity is that it provides a quick and easy way for the end user to determine if the network has basic connectivity. Link-integrity frames are sent once per second, unless there is traffic on the wire, in which case a reduced number of frames may be sent.
Since power-wire and wireless systems allow users to share the same physical medium, they must have encryption at the link level to attain a degree of privacy equivalent to the phone line. This requires some user key configuration, which somewhat defeats the plug-and-play objective. On the other hand, the presumptive privacy of phone wires is not a substitute for true cryptographic security.
Once installed, home networks are likely to remain in place for many years, and as home network interfaces become embedded in appliances, it may become almost impossible to replace them. Thus, a good home-networking technology ideally has built into the current generation a plan for interoperability with future generations. The HomePNA architecture should be applicable for some time, however, since it was designed to scale from 1-
to 100-Mbps transmission speeds.
Jason Trachewsky is a principal engineer with Broadcom's Home Networking Business Unit. He has also chaired the HPNA specification committee. He can be reached at jat@broadcom.com.
