Description
This video demonstrates the parity update overhead on an SSD with 8 chips organized as RAID-5. We run a Zipf workload in which the size of all write requests is one page. We expect every write to generate a parity update, resulting in a parity update overhead of 0.5. As the simulation progresses, the parity update overhead drops from 0.5 to 0.32 . The reason for this drop is that the data pages are inherently colder than the parity pages that protect them. Thus, they are more likely to be valid and copied during garbage collection. As a result, parity pages are responsible for a smaller portion of internal writes than data pages, and their update overhead decreases as the write amplification increases. The space occupied by the data and parity of the hottest stripes decreases with each garbage collection invocation.