Interpreting a Correlation Analysis

With Correlation analysis, the user inserts a number that looks a likely candidate for some connection to the problem (a packet length, bus width, ECC block size, counter period, etc.). This "magic number" is then used to define the number of bins in the correlation graph before wrap-around happens. If there is a correlation of errors to the magic number, discrete peaks will be seen. If not, the distribution is flat.

1a. When doing Correlation analysis to the magic number eight, having all errors in one bin shows that the errors directly correlate to the divisor eight.

1b. If we did this same Correlation analysis, but looked for a correlation to a different number like seven, we would find that no correlation exists. We see this because there is no direct relationship to the new correlation factor, so all bins are equally filled.

2. Errors that show a strong correlation to some number (like eight) cannot be explained by serial bit error sources that are random in nature. For example, this result cannot be blamed on a serial link—it looks like a problem with a digital word/eight-bit wide bus, or a faulty pin on an IC.

3. 1632 bits (eight bits/byte x 204 bytes) is the number of bits in an MPEG packet. All errors in this example are happening at one location in the packet. This might be due to a bad cell in a memory chip.

1. An external Marker signal connected to a disk drive Index pulse gives a correlation analysis that relates to every rotation of the disk. If all errors were random, you would see a flat line—all locations on the disk had the same probability of error. Errors caused by motor commutator switching noise may appear like this.

2. The same type of analysis may show that the error rate is lower in one half of the disk rotation than the other half. This may be due to increased head-surface gap spacing for half of the rotation, due to platter distortion or mechanical modding.

3. Seeing a flat distribution of errors with respect to any Marker is the indication that no correlation of error exists with that signal. This usually means you are done with that signal—time to move on to any other correlations that might exist.

4. Seeing a spike in the correlation analysis is a great indicator that you have found something. You might have more than one spike. When writing a sector of data, correlating errors to the sector boundary is a natural analysis to do. If a spike is found, try re-writing the data to see if it happened during the write process or read process.