The 5,000 digital optical modules (DOMs) could have been phase locking; the technology exists. For instance, clock recovery based phase locking could have been implemented. The possibility that a single element of the phase locked system could bring the experiment to a halt motivated an alternate solution. Further, the possibility that the (unmeasured but expected) radiation of the collection of 5000 phase locked devices could have exceeded a radio frequency/EMI emission level also exists.
To avoid a certain level of system complexity, and to avoid coherent EMI, the decision was made to install free-running crystal oscillators in each DOM. Each DOM local clock drifts in a random and unpredictable way due to crystal aging and other more subtle effects. Consequently, the experiment clock system monitor must track the frequency and phase of each DOM local clock oscillator using specialized software and hardware. To facilitate this tracking, timing signals coexist with and are integrated into the telecommunications signaling.
The frequency/phase tracking update rate depends on the short term stability of the crystal oscillator in each of the DOMs. If the short term stability of an oscillator falls below a certain limit, no update rate can be fast enough to provide the required system accuracy. That level is probably around 1E-8.
If one assumes that short term stability relates to relatively rapid but smooth frequency drift, then, one would be lead to believe that a short term stability of 1E-9 would be adequate to yield a 1 ns accuracy at a correction load of about once per second. Unfortunately, short term frequency fluctuations are often abrupt, 'hops' from one frequency to another, occurring at uncorrelated intervals, of uncorrelated magnitude. The hopping arises from the mechanical quality of the crystal (how good is the quartz purity and crystal freedom from defects), the 'cut', the Q of the resonator, the stresses of the crystal mount, crystal preparation and plating, and from particles and molecules that move around the surface of the crystal or are adsorbed and desorbed from the surface (preparation), to name a few. So, the short term stability of the oscillators in the DOMs must be somewhat better than 1E-9.
High quality, reference grade (typically oven controlled) oscillators deliver this level of performance. SC cut crystals are better than AT cut crystals. Their data sheets specify the Allan Variance as a measure of the short term stability. These reference oscillators typically consume tens to hundreds of mW of power. They exceed the power budget for the DOM. Other grades of oscillators include clock oscillators for digital applications, and telecommunications grade oscillators for products like cellular phones and modems.
Manufacturers of inexpensive communications grade oscillators seldom specify the Allan Variance. The nearest they come to a relevant specification is the SSB phase noise. These products can meet the 5 mW IceCube target.
If you are not familiar with Allan Variance, see http://www.allanstime.com/AllanVariance/ for a discussion. Additional references are also listed there. The Allen Variance data set consists of high resolution frequency or phase measurements taken over an extended period of time, and without dead-time between measurements. The nature of the analysis is such that the result should be independent of frequency and measurement interval.
In 1997 for AMANDA, and now for IceCube, the measurement data set consists of frequency measurements taken at exact intervals without dead time between measurements. Two Hewlett Packard model 52131A (twelve digit) frequency counters, gated alternately, produce a sequence of measurements. These counters do mathematics on the internal measurements, so can deliver a resolution of 0.0001 Hz for a 20 MHz signal with a gate of 5 seconds. The internal frequency reference of the counters (even the optional high stability one) is not good enough to make the measurements. Instead, a Stanford Research Systems model FS-700 Loran-C based frequency standard provides the 10 MHz reference. Our FS-700 has the optional low phase noise, SC cut, crystal standard installed in it. The specifications page indicates that the Allan Variance of the unit is about 5E-12. The programmable square wave frequency output of the FS-700 can provides convenient gating periods for the frequency counters. A LabView program controls the sequencing and read-out of the frequency counters. Care is taken that both counters receive their reference clock and gate with the same phase (equal length reference clock cables and equal length gate input cables).
The measurement limit of the system can be discovered by measuring some product that is better than the measurement system is. Our 'golden standard' is a model PRS-10 Rubidium oscillator module, also manufactured by Stanford Research Systems. Measurements of the Rubidium oscillator, when analyzed using the standard software package, gave the lowest Allan Variance measured with the system. (about 2E-11). see note
The package removes outliers based on the assumption that data
collection hardware can sometimes deliver erroneous data, and that
occasional large jumps in frequency are to be expected.
The package removes exponential drift usually associated with warming or cooling to equilibrium temperature.
The package removes linear drift (aging?).
The package removes quadratic drift (?). (I don't have a real world justification for that one, unless it is also thermal.)
The validity of this tampering with the data may be subject to argument. The 'corrections' obviously 'improve' the result. Do they hide some important feature or flaw of the product under test? On the other hand, few outliers are discarded, and the corrections are small, the result should be reliable/adequate for product comparison, or for acceptance testing of the chosen product.
Toyocom TCO-999, 100 samples. Samples ordered by the U of Wisconsin IceCube project office. Tests were performed in the fall of 2002. The population is similar to 50 oscillators measured several years ago, if not a tad better...
One sample of the TCO-999 failed during testing.... However,
by a wide margin, we haven't found anything with as good an Allan
variance for a comparable price...
Recent graphical results
benefit from better measurement technique, and more time to
accumulate samples. The 'best choice' for IceCube will be the one
with the best combination of performance and reliability. Linear
drifts obviously muck up the graph. Corrected recent graphical results,
and in an expanded scale, allow
one to get a better, but bussier, look at what's really
happening. In the end, great performance plus pedigree won out
over somewhat better performance with low cost.
Event - A collection of hits, which when analyzed, yield the identity and trajectory of a charged particle traveling through the detector array.
PMT - Photomultiplier tube, an exquisitely sensitive light detector capable of detecting single photons in a certain color band with better than 50% efficiency (i.e. at least half the time).
EMI - Electromagnetic interference. Unwanted radio emissions from electronics. Electronics devices may produce or suffer the effects of EMI, or both. Federal and international limits exist for the generation of EMI. The NSF specifies a requirements document.
SSB - Single Sideband. A variant of amplitude
modulation where only the modulation sideband energy is transmitted,
not the carrier.
The setting of the counter input trigger threshold affected this
difference. Correcting by level shifting may not, however, be
justified. The predictability of the error suggests that a simple
numerical correction could be used to improve