No clear SMART correlation with fail-slow NVMe SSDs: 4,584 of 779,978 drives, 10 later in the failure tickets

The 1.41% figure is a model average, not the share of drives scored in Table 7. The abstract says that on average 1.41% of NVMe SSDs are fail-slow within four-month monitoring, which is 6.05 times the HDD figure. Section 5.2 then ties that 1.41% to the 5-minute requirement and writes the multiple as 1.41% to 0.20%. At 60 minutes it writes 0.52% to 0.01%, 51 times. The reading that reproduces those two multiples is the excess over the HDD average, not the quotient: 7.05 and 52 are not in the paper. Table 7, under the same 5-minute bar, counts 4,584 slow drives out of 779,978 and prints that margin as 0.59%. The identification prose says around 5,000 fail-slow NVMe SSDs. Dong et al. restate the 1.41% sentence, place over one million NVMe SSDs in parentheses in that same sentence, and cite Lu et al. That paper does not restate the SMART result or the 0.22% transition. It is a citation, not a second measurement. Event frequency follows the same excess pattern: 48.00 versus 2.45 events per 1,000 drives per hour at 5 minutes, printed as 18.59 times, and 5.56 versus 0.08 at 60 minutes, printed as 68.50 times.

Fail-slow here is a write-latency fault, and the signal that catches it is a peer comparison, not SMART. Section 5.1 marks a drive suspicious when the 3-hour median write latency is above the third quartile plus two interquartile ranges. The paper calls a drive over that bar higher than the top 0.04% of latency variances in the cluster. A drive whose IOPS or throughput is also over the bar is dropped, so heavy traffic is not counted. The suspicious drive still needs one event at least twice as slow as the median of itself and its 11 intra-node peers, held for more than 20, 60, 120, or 240 records of 15 seconds (the 5, 15, 30, and 60 minute bars). The daemon records only those averages, and only from 9 PM to 12 AM. The log can prove that, inside that window, the median sat over the bar and a span cleared the 2 times peer rule. It cannot prove the drive was slow in the other 21 hours, and it never sees the drives the heavy-traffic rule dropped or the slowdowns that miss the 2 times bar or the minimum length. Section 6 says fail-slow has no clear oracle, the thresholds are empirical, and the strict bar may underestimate the impact. The paper expects the annual rate to be higher than this four-month, three-hours-a-day figure, and it does not print that annual rate. The same table's NVMe average event latency is 174.40, 155.17, 159.02, and 149.58 microseconds at the four bars. The paper calls that around 160 microseconds, SATA-SSD-level latency, and says the top 1% of events in several models reached around 22 milliseconds. Those are measured write latencies, not a bus rate.

The signal that does not catch this fault is SMART. Finding 9 says SMART attributes only exhibit negligible correlation with fail-slow metrics. On the last day of the four-month window the authors grouped drives by model, age, program/erase cycle, and workload, labeled each drive slow or not-slow, and ran a Spearman rank correlation. Under the 5-minute bar that is 40 groups. None showed a clear correlation with any SMART attribute. The paper names Critical Warning, P/E error, and CRC error as examples, and it does not print a coefficient for any attribute. The same null result held at 15, 30, and 60 minutes, and when slow meant more than one event. The artifact appendix says Findings 1-8 and 10 can be checked from the released dataset. Finding 9 is not on that list. The released performance logs cover around 97,000 NVMe SSDs and 141,000 SATA HDDs, not the half-million drives in the iostat study.

A later failure ticket also misses almost all of these drives inside the window the paper has. Table 7's replaced column is drives that appear in the failure tickets. Ten drives are in both the slow row and that column. Section 5.4 calls that around 0.22% of the 4,584 slow drives and around 0.23% of the 4,439 replaced drives, and the cell is printed as <0.01% of the 779,978 drives scored. The abstract's 0.22% is that slow-drive share, not a fleet rate. Mean and median gaps are 73 and 67 days. The latest tickets are about five months older than the last recorded fail-slow event. Finding 10 says the transition was rarely observed, at least not within about five months. That is not a claim that fail-slow never becomes fail-stop. The ticket shows a later failure record for those 10 drives. It does not show that the other 4,574 stayed healthy, and it does not show that the slowdown caused the ticket. Separately, the paper reports vendor repair results for the 100 slowest SSDs, around the top 2% of identified slow drives, average event latency 4.4 milliseconds: 33 bad capacitors, 46 bad chips, and the rest unclear. Those are the vendors' findings as the paper reports them, not a rate for the 4,584.