I've taken a break from fixing my server, and taken a look at MDWS for the new ECI G.INP lines.
All looks reasonably well... The retransmission counters show it to be being used, and in every case, it has reduced CRCs to, or near zero and ESs to zero. There is a variety of impacts on FECs, depending on how the line was behaving before activation
For the two lines that had DLM intervention before, skyECI and MikeZ, the FEC rate reduced considerably. I presume the deactivation of interleaving means that the FEC process became less effective, and that more failures fell through. However, those extra failures are now mopped up successfully by the retransmission process.
The two lines that had no DLM intervention, S.Step and jamesfoley, both had no FECs (no surprise if the FEC process is deactivated, as ought to be standard), and low/no CRC and ES counts before activation. After retransmission was turned on, alongside FEC activation, jamesfoley's line starts up with a low level of FECs, a low level of retransmission, and the loss of all CRCs and ESs. S.Step's line is better, as it starts and ends with no CRCs and no FECs, and show a low level of retransmission happening. Strangely, it counted a non-zero ES level, beforehand, even with zero CRCs. Of note for S.Step is the sudden appearance of prolonged spikes of upstream FECs.
That leaves two further lines, Simon194 and daveesh1. These two lines do not show DLM intervention beforehand, with a zero INP value and no interleaving. However, both lines showed FEC counts even before G.INP activation, suggesting that the modems had chosen to turn on FEC anyway. In both cases, activation of G.INP results in FECs continuing, and both CRCs and ESs dropping to zero, alongside some use of retransmission.
This thread contains an amount of speculation about why the downstream speed has changed in the way it has, some up some down. I have no magic answers, but I suspect part of the answer comes down to the relative changes to the FEC settings, and the relative amount of bandwidth being "stolen" for the FEC parity overhead. We'd need to see a lot more of the before and after raw framing data to figure that out.
We can do a little bit of analysis, though.
- First, we can see that DLM has used 4 different INP values - 46, 47, 49, 50. Presumably ordered from best to worst.
- jamesfoley's line has INP=46, and gets FEC settings of R=6, N=185. This means 3% of the line's bandwidth is now being used for this small amount of FEC alongside retransmission.
- skyECI/skeynewb's line has INP=50, and gets FEC settings of R=12, N=240. This means 5% of the line's bandwidth is gobbled up by overhead.
With a sample size of two, it is hard to reach proper conclusions, but it looks like higher INP values make for higher FEC overheads.
@BE
Skynewb's retransmission usage is higher than the others, and has a few more spikes, but doesn't appear that heavy. I'm not sure how much heavier it would need to be to start registering LEFTRS.