It is known that applying sophisticated disk scheduling algorithms such as SATF at upper levels in cache hierarchy is a fruitless enterprise. However, we have demonstrated that CSCAN can be profitably applied even at upper levels of memory hierarchy for effectively improving throughput. As the size of NVS grows, for a random write workload, the throughput delivered by CSCAN seems to grow logarithmically for single disks, RAID-10, and RAID-5.
CSCAN exploits spatial locality and is extremely effective in cache-insensitive storage configurations. However, it does not perform as well in cache-sensitive storage configurations, where it loses to LRW that exploits temporal locality. Since, one cannot a priori dictate/assume either a cache-sensitive or a cache-insensitive scenario, there is a strong need for an algorithm that works well in both regimes. We have proposed WOW which effectively combines CLOCK (an approximation to LRW) and CSCAN to exploit both temporal locality and spatial locality.
We have demonstrated that WOW convincingly outperforms CSCAN and LRW in various realistic scenarios using a widely accepted benchmark workload. WOW is extremely simple-to-implement, and is ideally suited for storage controllers and for most operating systems. WOW fundamentally increases the capacity of a storage system to perform more writes while minimizing the impact on any concurrent reads.