舰队闪光.pdf

1、Vineet Parekh,MetaMinh Hoang,MetaFlash in FleetFlash in FleetVineet Parekh,MetaSTORAGELatency MonitoringPrevious Limitations:Host-level checks missed extreme tail latencies(P99.999999+)due to their transient nature.Challenge:Intermittent SSD failures made debugging and data collection difficult for

2、vendors.Solution Latency Monitoring:Enables precise detection of tail latencies and stalls,improving visibility and SSD-level diagnostics.Thousands Of Drives EnabledSSDs with Latency Monitoring activeHow Latency Monitoring WorksLatency Buckets:I/O completion times are categorized into configurable l

3、atency ranges.Operation Tracking:Monitors read,write,and TRIM operations across these latency thresholds.Dual Counters:Uses active(short-term)and static(long-term)counters for real-time and historical insights.ClassificationEach I/O operation is sorted into one of four latency buckets based on compl

4、etion timeThresholdsHost-defined boundaries(e.g.,100ms,250ms,500ms,1000ms)determine bucket assignmentPersistenceData persists across resets and power cycles for long-term analysisLatency Bucket ConfigurationBucketLatency RangeExample ThresholdBucket 0 Threshold A100msBucket 1 A and B250msBucket 2 B

5、and C500msBucket 3 C and D1000msThresholds A through D are configured by the host system and define the latency boundaries for different operation types(read,write,and TRIM commands).This bucketing approach allows for granular analysis of performance degradation patterns.Core Components:Active Count

6、ersActive Bucket CounterTrack I/O latency in the current window;reset periodically(e.g.,every 7 days).Ideal for short-term diagnosis and recent latency spikes.Active Measured LatencyCaptures mean latency(ms)of I/Os within each bucket,indicating severity.Returns zero if unsupported or no matching I/O