Relaxation Oscillation in Gigabit Ethernet and Fibre Channel
Relaxation oscillation may not have been a serious problem for low-speed optical links, but it can be a major issue for gigabit links such as Gigabit Ethernet or Fibre Channel. This
article details techniques for measuring relaxation oscillation and for determining its impact on the overall system.
By Clark Foley
Relaxation oscillation (RO) is a fundamental characteristic of Fabry-Perot (FP) lasers. It occurs when the laser is switched from an off state near or below threshold, to an on state well above threshold. The oscillation can vary in frequency, amplitude, and duration. In early short-wavelength Fibre Channel laser transmitters, manufacturers used lasers that
were made for CD audio players. When data rates were 266 Mbps, it was easy to sort lasers intended for CD players for desirable features. Two features these early lasers exhibited were an RO of over 1 GHz and a comparative cost-effectiveness. They were also band-limited in such a way that the RO frequency was well out of band and rarely manifested itself enough to degrade performance. As Fibre Channel moved to 1.063 GHz, it became necessary to select the lasers more carefully, eventually creating special
fabrication runs to get an acceptable RO for this higher bit rate. For 1.063 Gbps, the RO needs to be more than 2.5 GHz and must decay quickly. With a few exceptions, manufacturers of Fibre Channel transceivers struggled with the part-to-part variations resulting from different suppliers or fabrication runs.
Depending on the lot date, supplier, and operating point of the device, variations in final test output can continue to require excessive reworking and the occasional engineering fire drill.
It’s an expensive operation. With the deployment of Fibre Channel-based terabit disk storage arrays and Gigabit Ethernet, the necessary supply of short-wavelength (780 to 850 nm) optical transceivers can’t be attained without higher production yields. The need to reduce part-to-part variation is pushing suppliers toward the development of an alternative — the short-wavelength vertical cavity surface-emitting laser (VCSEL). Recent success with VCSELs has led some suppliers to phase out the FP
lasers where possible. Furthermore, VCSELs have not yet been mass-produced for long-wavelength (1,310-nm) transmission. This article describes testing methods used to evaluate and predict the impact of RO in Fibre Channel and Gigabit Ethernet applications.
RO in the time domain
A good first step in understanding how RO will influence the quality of a transmission signal is to examine the RO directly in the time domain.
Figure 1
shows the output signal
for a short-wave gigabit interface converter (GBIC). The signal was produced using a high-bandwidth (8-GHz), broad-wavelength (700 to 1650 nm) optical-to-electrical (OE) converter, with a high-bandwidth oscilloscope. Stimulating the GBIC using the Fibre Channel K28.7 character (five 1s followed by five 0s) permits signal averaging and smoothing to reduce noise.
Manual adjustments of cursors or markers measure features of the signal, such as the frequency or the period of the RO’s initial peaks.
(Note that the overshoot is 200%.)
Measuring the RO frequency
The time between the first two peaks of the RO is easy to measure. It is also easy to manually adjust cursors or markers to measure the frequency or period of the first two peaks of the RO. The cursors in
Figure 1
report that the time interval is 420 ps and the initial RO frequency is about 2.4 GHz. If the RO persists for several cycles, the neighboring peaks are also measured. Manually
adjusting cursors can quickly become tedious if it is necessary to evaluate many parts at different temperatures, or when done in the production area where hundreds of parts must be tested each day. It shouldn’t be difficult to replace the skilled technician or engineer with an automated alternative. Some oscilloscopes can search for peaks, while others can search for level crossings, or have built-in digital signal processing to manipulate waveforms and extract data.
There is a method that
can be used for peaks, troughs, and crossings that eliminates the need to estimate where to set the crossing level. This method can be implemented with either the high-bandwidth, 8-GHz OE converter or (as shown in
Figure 2
) a 2.3-GHz multirate optical reference receiver. This approach is a reliable way to automate the RO frequency measurement, and makes use of the built-in waveform computational capability of the oscilloscope. It’s not hard to measure the period or
frequency of the bottom trace.
In
Figure 2
, the top trace is captured using the full bandwidth (2.3-GHz) selection of the optical reference receiver. Although significantly lower in bandwidth than the 8-GHz OE converter, the optical reference receiver is still able to capture the RO. The RO frequency measurement is automated using the built-in waveform computational capability of the oscilloscope. After acquiring the waveform, the first derivative (middle trace) is performed
to eliminate the pulse and set the location of the RO peaks and troughs at the zero-crossings. If desired, the second derivative can be taken to extract the zero-crossings of the RO. The middle trace is further processed by applying the Signum function. Signum sets all positive values to 1 and all negative values to -1. This is essentially a comparator function that greatly simplifies the automatic searching for level crossings on a waveform. Measuring the period or frequency of each cycle of the bottom
trace is easy. The high-frequency transitions before the acquired trace’s initial step are due to the noise around the middle trace’s zero value. Searching for the rising edge of the top trace will yield the search starting point. From there, various search methods can be employed to find the clean transitions of the bottom waveform. Figure 2 shows the left measurement delimiter, excluding the information prior to the initial step of the top trace. The automatic frequency and period measurements
find the first complete cycle to the right of the left measurement limit. The results of
Figure 2
agree with those acquired using the 8-GHz OE converter shown in
Figure 1
. The RO frequency is 2.38 GHz and the period is 420 ps.
Impact of RO on the eye pattern
With a clear view of the RO in the time domain, the RO will have a significant interaction with the data transmission. The oscillation undershooting below the decision
level (50% of nominal pulse height) could cause bit errors. Also, when bit transitions collide with the RO ringing, data-dependent jitter is created. Depending on where in the RO’s cycle the data transition occurs, the falling edge will be displaced earlier or later than the ideal position. If not controlled or suppressed, the RO will create an undesirable increase in horizontal and vertical eye closure. For this particular GBIC sample, the RO amplitude is extreme. The overshoot is 200%, and the undershoot
dives to the logic-zero level. Fortunately, the application calls for a bandwidth that is 0.75 times the bit rate and a rejection following the 4th-order Bessel-Thompson response. For Fibre Channel, the specified frequency at which the power is -3 dB is 797 MHz, and for Gigabit Ethernet it is 938 MHz.
Will a receiver with one or the other bandwidth be able to reject this wild RO? The transceiver designer will have to decide. Even if more bandwidth is available to the designer, he knows that more
is not always better.
The eye pattern mask test is one way to evaluate transmitters. This test is part of a conformance test stipulated by governing standards such as IEEE-802.3z for Gigabit Ethernet and ANSI X3T11 for Fibre Channel. Although these standards do not state what the time domain response of the receiver should be, they do stipulate how the transmitter should behave in a standardized measurement condition. These standards prescribe that the transmitter should be evaluated using a mask (a
group of violation zones through which the signal must not pass). It is further required that the mask be applied only if the combination of the oscilloscope and OE converter complies with the frequency response characteristics of the 4th-order Bessel-Thompson filter. The compliant combination of hardware and software is called an optical reference receiver. The reference receiver used in the examples in this article is a multirate optical reference receiver, compliant for use with Gigabit Ethernet 1.25
Gbps, Fibre Channel 1.063 Gbps, and OC48/STM16 2.488 Gbps (full bandwidth selection). This reference receiver can function over the wavelength range of 700 to 1,650 nm, making it ideal for measuring Gigabit Ethernet and Fibre Channel long-wavelength and short-wavelength transmitters. The full bandwidth selection permits investigation of the RO. This means that one tool and its associated methods can be applied to devices of both wavelengths.
Figure 3
illustrates one way
of evaluating the effect of the RO on the eye pattern. A simple stress test is performed by stimulating the sample GBIC with a special bit sequence. In this test, the logic level remains low for the maximum allowed time, followed by the shortest possible time it can remain high. This constitutes what is often called an isolated-1 sequence (1111000010). By the band-limited nature of the test, the isolated-1 pulse doesn’t have enough time to completely reach 100% of the logic-1 level before it
transitions back to the logic-0 level. Consequently, with the RO superimposed, it poses a challenge to passing the mask test. To estimate the behavior of the transmitter with such a bit sequence, the full bandwidth (8-GHz) signal is acquired using the OE converter. This is the upper trace in
Figure 3
. With the help of the oscilloscope’s internal DSP filtering ability, the signal is mathematically filtered to approximate the Gigabit Ethernet, 4th-order Bessel-Thompson response of
the optical reference receiver. The result (the lower trace in
Figure 3
) predicts that the short isolated pulse will be unable to remain above 80% of the logic-1 level, which is needed to pass the mask test. The result also shows that a significant amount of the RO will be retained.
The actual outcome is dramatically illustrated using an eye pattern, shown in
Figure 4a
and
Figure 4b.
Because transmitter masks are
defined to evaluate deterministic waveform effects, random noise and jitter have been eliminated during the acquisition of these eye patterns. This provides a clear view of the paths various bit sequences take in forming the eye pattern. In Figure 4a, the optical reference receiver is used as the Gigabit Ethernet reference receiver. In Figure 4b, it is used as the Fibre Channel reference receiver. It’s clear that the sample GBIC is failing the Gigabit Ethernet mask test as predicted in
Figure 3
. The excessive ringing results in an intolerable vertical eye closure. Using the reference receiver set for Fibre Channel shows that the RO suppression is good enough to pass.
RO: friend or foe?
The RO is often relied upon for a fast-rising edge. Unfortunately, its unpredictability in frequency will make a marginal or poor pulse response look good in one batch of parts and terrible in another. There are, however, many ways to inspect, predict, and evaluate the
impact of the RO on the transmitter eye (as shown by the conformance mask test). Controlling RO is essential to maintaining link margins for optimum performance.
Clark Foley is a design engineer for the measurement and accessories product line at Tektronix, Inc. He has a BSEE from Arizona State University in Tempe, AZ. He can be reached at
clark.foley@tek.com
.
Return to
Table of Contents
Puzzled by a network processing design issue?
