I also suggest that you will try to use %PARALLEL. sometimes it helps.

We have experienced that in some (very heavy queries) a good option is in some cases is to have a SP (stored procedure) that will get the query request parameters and run it in parallel "segments" using the build in "queue manager"

Hello Peter,

Welcome back to the community.

I also have almost ~30 years (since 1991) experience on Intersystems technology starting with some PDP11, VAX & Alpha machines (anyone else misses the VMS OS like I do?) with DSM, the first PCs with MSM and then (1999) came Cache that become IRIS.. very long, interesting and challenging way !

I do remember at 1992 customers with 50-100 users running MSM on a 286 machines. and it was fast (!) At that times, where disks were slow, small & expensive, developers used to define the globals in a very compacted & normalized way, so applications flew on that platform.

Today, when disks (as Rebert said) are without limits, fast & relatively cheap, some "new" developers tend to disparage with a correct design of a DB, which on a big databases of few TBs it is noticeable.
You can get a significant improvement when the are been optimized (this is part of my job in the past few years). and this is not just the data structure, but also in correct way of optimized coding.

In my opinion, the multi-level tree of blocks that hold a global, is limited by the DB itself, which is limited by the OS file system. With global mapping, and sharding (new technology) - even a DB in a size of PETA bytes coukld easily hold them, and be a very efficient, like this technology was forever.

Here is a test for running all those codes. 
It looks like that the 1st solution (with $translate) is the fastest.