Hi,
Because of your writes being faster ( because the writes go to disk cache instead of directly to disk ) the background writes caused by exceeding the LRU_MAX_DIRTY threshold will also be faster. Therefore buffers can be liberated in memory faster. This then can have the effect buffers have to be read from disk again when they have to be modified, resulting in a worse write-cache ratio.
Make sure you reinitialize the figures ( onstat -z) periodically to be able to compare them .
Best regards
Koen