Please note that these are very preliminary results and do not necessarily have much bearing on the expected performance of fasttime.
As an initial test of the metrics and justification for the project, this test was run using the currently implemented "dummy" implementation. The clock is calibrated to the system clock in 1 second before any measurements are taken, and is subsequently not adjusted (allowed to run free). The resulting error in offset and rate can be attributed to any of variation in the TSC clock (the clock being measured), variation in the system clock (which is used as a reference) or latency in the system (increasing the measurement dispersion).
The test was run on a 900MHz G3 iBook running OS X 10.4 over a period of 12 hours. During the experiment the computer was under virtually no load (it wasn't touched and was serving no network requests). The computer is indoors and was probably subject to temperatures between 10 and 20 degrees Celsius.
| Offset | Rate error | Allan deviation |
|---|---|---|
|
|
|
|
The same test was run, again over 12 hours, on a P3 800MHz running Linux 2.6 Fedora Core 2. Time on the machine graciously provided by Internet2.
| Offset | Rate error | Allan deviation |
|---|---|---|
|
|
|
Due to the large discontinuities in the above experiment much of the detail is lost. The first 2 hours of the same experiment is plotted below; no discontinuity exists in this period, so a more interesting comparison can be made with the Mac, and forthcoming fasttime results.
| Offset | Rate error | Allan deviation |
|---|---|---|
|
|
|