Decreasing CF memory is slow

While working with PureScale, there may be a need to create room for new DB in the system. Once this happens we need to reconfigure CF memory, while keeping system up and running.

We are concerned regarding CF_MEM_SZ , CF_GBP_SZ and CF_DB_MEM_SZ, which we can not decrease online even though those parameters are configurable online.

Sample command that results in timeout:
db2 update db cfg for prdind using CF_GBP_SZ 5243120 immediate
SQL2046N The configuration parameter named "Cf_gbp_sz" could not be updated
because the operation timed out.

Note that timeout above happens as the value we try to decrease is too large. If we try to reduce in small steps, command works fine.

Based on our analysis, it is difficult to shrink a structure. Shrinking requires moving memory around depending on many factors like de-fragmentation, workload (latch contention), potentially more factors. This is why the automatic CF memory tuner always shrinks in very small increments like 1MB. There is no guarantee that a structure can be shrieked to any value within a given time limit. Especially GBP because its always in use during a workload.

To be able to shrink a structure more effectively, we have decreased in small chunks. This is time consuming, and if we need to release many gigabytes, it can easily take many hours with current behaviour.

We also considered castout that can be configured to be more effective, so that pages can be freed in the GBP so that they can be re-used. This configuration is related to GBP castout (page_age_trgt_gcr), however this parameter has performance impact, and difficult to modify in PROD.

Yet another alternative is to consider offline change for CF memory, to create room for restore test faster. However this does not maintain HA as we need to bring CF offline.

Can we get a design change, so CF memory can be tuned more effectively, and entire PureScale solution is more robust and user friendly?