Looking at the stats for all the lines not just the repaired ones, I see a few ES and SES in the recent 15 min periods. Line 1 with it’s crazy upstream of 666k was achieving this at an upstream SNRM of only 2.5 dB not 6dB, so that explains the silly high figure and suggests my analysis was right: perhaps retraining during an artificially quiet time and then noise rockets up, but anyway a very high noise level now which presumably the modem cannot cope with therefore causing corruption and TCP retx which slows throughput down to a crawl in a TCP throughput based tester. This shows the advantages and disadvantages of such a tester design: it is real-world, but it isn’t measuring the line - it’s measuring the particular protocol and its implementation. Testers should test with both methods and report two numbers, and should also report corruption and packets dropped in transit.
So I forced a retrain on each of the lines and things looked rather a lot more normal; downstream speeds lifted to where they should be. Post forced retrain - Live sync rates:
#1: down 2990 kbps, up 560 kbps
#2: down 2880 kbps, up 547 kbps
#3: down 2998 kbps, up 512 kbps
#4: down 2949 kbps, up 502 kbps
Note line 3 upstream still very much cured, has not yet gone back to being sickly as before. And the upstream rates are fairly close, which is always good.
Firebrick current upstream rate limiters' IP PDU tx rates (egress speeds), in-force right now ::
#1: 477948 bps
#2: 466852 bps
#3: 436981 bps
#4: 428446 bps
Total combined rate: 1.810227 Mbps
Fractional speed contributions:
#1: 26.403% [████████████ ‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒]
#2: 25.790% [████████████ ‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒]
#3: 24.140% [███████████ ‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒]
#4: 23.668% [███████████ ‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒‒]
--
Obtained from live-querying the Firebrick.
Unfortunately the speedtest2.aa.net.uk upstream throughput test was still poor at 1.31 Mbps, so nothing seems to make sense. At one time it was 200k better, and so it is not the case that the bonding simply is inherently not that good; also the downstream bonding is very efficient.