Observing greater than ~11% drop in transaction throughput between 1 collection vs 1000 collections with durability level greater than majority

Observing ~18 to 20% drop in kv_transactions throughput between single collection vs 1000 collections .

http://showfast.sc.couchbase.com/#/timeline/Linux/kv/transaction/Throughput

Comparison tests - 

Avg Throughput (trans/sec), Workload T, 4 nodes, Majority_Persist_Active, 12 vCPU, s=1 c=1000, worker_instance=1

vs

Avg Throughput (trans/sec), Workload T, 4 nodes, Majority_Persist_Active, 12 vCPU, s=1 c=1, worker_instance=1

 
Test 

stats comparison

http://cbmonitor.sc.couchbase.com/reports/html/?snapshot=hercules_700-4617_access_a404&label=1000collection&snapshot=hercules_700-4617_access_6c6d&label=1collection

Avg cpu utilization in case of 1000 collections is significantly higher compared to single collection .

                                      java client side cpu comparison

It's known that we have seen about ~6% drop in kv throughput with large number of collections (mostly to sanitize incoming collections in each operation) MB-39594.

However trend seen in case of kv_collections seem to be alarming considering degradation is quite indeterministic across different durability levels .

*I will share consolidated comparison across different durability levels soon .