What could be a reason of massive cache flush?

[Павел Суслов]

Hi!

I have a cluster with 4 shards. Hardware is identical on each shards. Cluster has intensive inserts permanently (approx. 1-2K/sec). And a lot of queries (up to 6K/sec).

Sometimes almost all data in cache flushes. At this moment I see significant IO performance impact because of reading from HDD. Every day I create new database/collection for new data. Every collection is about 70G. It can be used simultaneously up to 7 collections (for queries). For inserts only 2 collections can be used. All queries uses indexes. More than a half of such queries has no results (checks for duplicate values etc). Sharding key of each collection is a string, containing hash (md5).

What could be a reason of cache flushing. And how can I improve the performance? Solving the problem of cache flushing would be a good point.

Thanks in advance