What to do about Fiber Skew? White Paper July 2008 www.commscope.com
A number of years ago there was a good deal of discussion about the problem of fiber skew in high speed optical links, but the issue has taken a back seat as laser optimized fibers became available to extend link distances for 1 and 10 Gb Ethernet systems serially. The development of 40-100 Gb Ethernet links using parallel transmission has once again brought up skew as a possible concern. This article will show that we can continue to install standard optical fiber trunk cables used today without the need to worry about their performance in future parallel optic applications. Skew likely won t be a problem for loose tube MPO trunk cables. Skew is strictly an issue for parallel data links, well-known in the electronics world ever since logic circuits containing more than one bit were invented. Most CPU based systems today transfer data and instructions in the form of digital words or Bytes. For instance the letter F at the start of this sentence is represented by the binary code 01000110 and the computer being used to write this article sends this letter back and forth to the memory via a parallel data bus. It uses eight electrical paths formed into the circuit board and sends a 0 on the first path, a 1 on the second path, a 1 on the third path and so on, all at the same time. It would then send the letter o and then r and so on until all of the data was transmitted. As one can see why this is obviously a fast way to transmit data, the problem arises when the bits of a letter arrive at different times and get out of sync. Suddenly a 01000110 becomes something else and you get errors. Electrical engineers know this problem all too well as they are now normally working with data busses that are 32, 64, or even 128 bits wide and at very high data rates. They have to pay special attention to not only the difference in path length of the wires but also the electrical characteristics of those paths as these can affect skew. To a certain degree they know there will always be some skew, but as long as the difference in arrival time for the earliest to latest bit within the Byte or Word does not exceed the set-up and hold times about the clock edge, errors due to skew will not occur. Greater design difficulty arises as the data rates get higher and the clock cycle gets shorter thus allowing less time for skew. Skew in optical interconnects is very similar to that seen in electrical systems, but can be even more complex to deal with physically due to the higher speeds of the data links as well as the longer distances the data has to travel. As mentioned previously, the optics industry started to get concerned about skew in optical systems but somewhat sidestepped the issue years ago with the overall slowdown in the industry as well as the refinement of Laser Optimized fiber. Using better fiber the industry was able to Serialize or take all of those parallel paths and squeeze them into a single optical path thus avoiding parallel optics for a while. www.commscope.com 2
SERialization/DE-Serialization (SERDES) is the process of taking a parallel data bus and transmitting each of the bits in time-sequenced order, one at a time, over a single path. This is typically accomplished via a SERDES chipset where the Serializer is located between the parallel data bus and the Optical Transmitter. Obviously the optical link has to be operating n times the speed of the parallel bus, where n is the number of parallel paths in the parallel bus, or it would become a bottleneck. At the other end of the link, once the optical receiver converts the photons back into electrons, the de-serializer reconstructs the parallel electrical signal and sends it on its way. As far as each individual optical link is concerned, it could care less about any other optical link it is bundled with as all of the bits it cares about are transmitted by themselves on only one fiber. This method works fine as long as the data rates do not exceed the capability of the optical link. But once you hit that limit, multiple optical paths used in parallel offer a low-cost way forward, just like with electrical busses. This technique bypasses the SERDES and just maps the data bus right onto the fibers. However, now the optical receiver must properly handle the effects of skew. The architects of the Ethernet protocol have devised a partial workaround to true parallel links via Link Aggregation. In this case one still relies on serial optical transmission at the physical layer, but utilizes LACP at layer 3 of the OSI model to make the system think it has only one link versus 4, 8, 10 etc. Data from a single flow (i.e. conversation) is assigned to traverse on only one of the aggregated links so that its packets are received in the same sequence that they were transmitted, and all of the routing overhead associated with this function must be handled by the CPU. www.commscope.com 3
As was stated previously, dealing with skew in the electronics is nothing new. Some fiber and cabling vendors have been trying to scare the industry into tighter skew specifications for the physical media because, as you can guess, they want to drive customers to their product. Skew in optical fiber can come from a number of different areas. One of the major sources of skew, at magnitude of roughly 13 ps/m, is caused by index of refraction variation from fiber to fiber. The index variation is limited by fiber standards that place upper and lower limits on the Numerical Aperture (NA) of the fiber. No matter how one packages the multiple fibers in a parallel link, be it in a loose tube, tight buffer, or ribbon cable, one will always have to either deal with this variation electronically or select matching fibers for each cable in a parallel link. Dealing with skew electronically involves creating a buffer or queue to store enough of the data from each of the lanes to allow realignment. The system must have enough intelligence to determine how to reorder the lanes. For example, it can take advantage of certain framing protocols that imbed location information. While there is some latency imparted into the link due to the buffers, at 40Gb or 100 Gb rates it will be far lower in cost than a comparable serial optical link. A second cause of skew in some optical links is physical length difference of the fibers in the cable. In this area ribbon fiber can help minimize the overall effect, but as will be described later, ribbon has its own issues. We estimate that the worst-case physical length difference in loose tube cable will impart approximately 25 ps/m of skew. As the system will be designed to accommodate the skew caused by the index variation of industry standard OM3 fiber, the same correction technique and circuitry can address this physical skew attribute. Both of the prior two contributors to skew are fairly constant in time thus allowing for close to a set and forget philosophy for correction. A third dynamic skew attribute that really only manifests itself in ribbon fiber or any tightly bound fiber cable structure is variation based on strain-optic effect. This dynamic skew will require correction techniques that can constantly shift and correct. The index of glass varies slightly when forces are imparted onto it. This additional index variation will cause a slightly different propagation velocity for fibers under stress for those that are not. We estimate that this dynamic effect can be up to 4.4 ps/m. This magnitude is by far well within the static amount that can be corrected electronically with the same circuitry needed for the previous skew contributors, but the dynamic nature of this skew has the added complexity of changing randomly in both magnitude and direction in time. Ribbon constructions can be particularly prone to the strain-optic effect. In many cases the fibers on the outside edges (fibers 1 and 12) see higher stress due to the fact that they are bonded in the flat linear ribbon structure and can see different stress as the cable is bent. The amount of this stress can vary over time as cables are disturbed or they exhibit thermal expansion/ contraction via heating and cooling. The dynamic nature of this skew imposes additional requirements on the skew-handling circuitry, which not only must account for the maximum skew, but also the maximum skew variation. This added complexity will add to circuit cost and potential latency. www.commscope.com 4
The lowest-cost implementation of 40 Gb (or 100 Gb) transceivers will consist of 4 (or 10) individual 10 Gb devices packaged into an array. While the use of array optics saves space, the true benefit comes from being able to treat the array as a single Physical Layer channel. The underlying electronics can deal with the skew so that the individual received 10 Gb lanes can be properly reordered to recreate the original 40 Gb (or 100 Gb) data stream. The IEEE 802.3ab Task Force is currently working on the 40 Gb and 100 Gb Ethernet standard, including de-skewing specifications, and expects it to be published by 2010. In May of 2008, CommScope proposed a worst-case skew model for parallel multimode channels, which was accepted by acclamation of the Task Force. De-skewing circuitry will be built to handle the worst-case skew of this model. This means that skew-controlled cabling will not be required to support the specified channels for 40 Gb or 100 Gb Ethernet. One way to think about skew-handling techniques is by analogy. Is it better to build cars that have the appropriate suspension to handle the typical variations found on the surface of roads, or to take the function out of the car and force our towns and cities to have hyper flat and smooth road surfaces? The answer is obvious. Likewise, given the ever-decreasing cost of transistors in integrated circuits, it is far better to build skew correction into the electronics versus creating some superficially attainable skew-controlled cable specification that really only drives up costs. The latter is a great idea for cable companies looking to make premium ribbon cable, but offers questionable benefit to the end customer. IEEE 802.3ba recognizes that the electronics-base approach ultimately provides the lower-cost, better performing solution. www.commscope.com Visit our Web site or contact your local CommScope representative for more information. 2011 CommScope, Inc. All rights reserved. All trademarks identified by or are registered trademarks or trademarks, respectively, of CommScope, Inc. This document is for planning purposes only and is not intended to modify or supplement any specifications or warranties relating to CommScope products or services. 07/11