We currently evict write cache buffers to the disk as soon as they are dropped onto the dirty list. This behavior limits our ability to exploit very large write cache buffer pools since the dirty list is always very short.
The write cache service should be modified to (A) defer eviction of write cache buffers to the disk as long as possible (without starving the application of buffers to install new writes) and (B) coalesce write cache buffers before they are evicted to the disk.
By deferring write cache eviction to the disk, the write cache buffers will be significantly more sparse as records will remain in RAM and have a much higher likelihood that they will be recycled rather than written to the disk. Further, by deferring evictions should be able to drive up the write throughput remaining 100% in RAM for data sets whose final allocation count is LTE the capacity of the write cache buffer pool in bytes.