Cache miss ratio increased in scenarios with a fixed working set and uniform access pattern

! This might be a test issue !

Many DGM tests follow these steps:

• Load data.
• Read a subset of data (e.g., 5%). This working set usually fits into the memory.
• Run the workload.

The probability of hitting the working set is explicitly defined (e.g., 95%). Both "hot" and "cold" items are accessed using uniform key generators. Those simple techniques allowed us to achieve a desired and constant cache miss ratio (e.g., 5%).

That no longer works in recent Vulcan builds (see attachment).

A more specific test:

• 4 nodes
• 1 bucket
• 1B items (~1KB)
• 15K ops/sec (90/10 R/W)

Graphs: http://bit.ly/2rdA22Q