How to clear Velox caches

[sumokiama]

We are running TPCDS 3TB tests on Gluten/Velox/S3/Centos8 and would like to clear all Glute/Velox caches between queries.

Running the same query repeateadly, the TPCDS elapsed time decreases for each run.
If we clear the system buffers between runs, there is no effect on elapsed time...it just gets shorter and shorter.

Note that we are running on a single server with a simple local Spark configuration.

We see that during a run the directories /tmp/blockmgr-* and /tmp/gluten* are created.
The gluten directory appears to be used for spilling which apparently not occurring as the directory is always empty.
The blockmgr directory contains a number of shuffle-* files.
Both directories are removed after run completion.

In addition to instructions on how to clear the caches, please advise on where the caches are located.