How much does a cache really write to flash?
A cache on an SSD writes more than its clients send it. The cache's own layout, the filesystem's journal and metadata, and the drive's garbage collection each add writes, and past a point the drive can no longer keep up. thrash.live simulates each of those layers for hours of drive time, so you can compare designs before building one.
Compare
The drive's flash
Every erase unit on each design's drive, filled to the share of its data still valid. Garbage collection must copy that valid data before it can erase a unit, so partly-full units are what make a drive slow down.
Results
Drag across a chart to zoom every chart to those hours; double-click to show the whole run.
More measures
Inside each design
What the device writes, by origin, over time; request latency percentiles; and the bytes each layer writes per byte the clients store. Each latency point's percentiles are over all the requests in its stretch of drive time, from a histogram with 8 buckets per power of two (within about 4.4%); where a point has too few requests for a percentile (1,000 or fewer for p99.9), it shows the largest, and a point with none breaks the line. The latency is the model's queueing delay: it has no per-request device variation, so its tail is backlog.