StarData has indeed much worse performance than AstronomicalData and a few more hickups.

1. Retrieving all entities, e.g., StarData[] takes a few minutes. However allNames=StarData["EntityCanonicalNames"] downloads almost immediately with a list of all star names that can be used to query again (this idiom needs to be found first, while the StarData[] is more obvious). There appear to be no integrity checks. Sometimes the list is corrupt, so it needs to be checked first.
2. Combined downloads, e.g., y = StarData[x, {"AlphanumericName", "DistanceFromSun"}] take forever, e.g., 6h+ for absolutely no reason, not even throttling should be that slow. This was doable ok with AstronomicalData. In this slow mode there also occurred spurious retrieval failures, e.g., { {"HIP109905", Missing["RetrievalFailure"]}}, where the object should clearly be available. Distances also have mixed units, e.g., kpc for the failed case above, instead of the default ly, requiring further manual cleanup.
3. StarData[allNames,"singleProperty"] appears to be running a bit faster, stepping first in units of 64 instead of 1, but sometimes throttling in steps down to 1. So if I want to sort for DistanceFromSun and retrieve data for a subset of e.g., 1000 stars, the fastest is to download the entity list of canonical names, then all the distances, then Join the two lists, takes ~50min, best case, possibly several hours. This is still much faster than the 6h+ with combined properties. In any case all ridiculously slow.
4. The entity queries appear to be somewhat unwieldy. Retrieval of the entity name list in no time shows that there is probably an issue with retrieving entities and the inferences necessary to retrieve more than one property.
5. In addition sometimes the kernel or all of MMA (11.0.1.0) dies, requiring several hours of download again..., so careful and cache your data locally once retrieved (Somehow AstronomicalData appeared to do more local caching. Could someone please confirm?). The old data caching infrastructure for curated data appears unfortunately gone, it may be a good idea to reestablish it. Even caching the whole database locally maybe an effective workaround for now.

In the current form StarData does not appear viable for work, only for single toy queries in Wolfram Alpha. The data amounts are small (e.g., 6 x 1000 Entities/values) and my notebook is under 1MB.

As a classic professional MMA user I would much prefer that those data issues are fixed than to have more tablet functionality and phone apps implemented. Can I have a choice using entities or just numerical values? Do I really need those units everywhere, buying one more inference layer?

(Remark: slowness is unlikely due to my machine or internet connection. The machine is a hexacore E5 workstation with 64GB ECC memory and Windows 8.1 providing a very robust high performance environment.)

Further input and hints for optimization are of course appreciated!

You can have that forms returned also with StarData directly, and the syntax is nice. Clearly better than AstronomicalData. It's the performance alone that ruins this.

There was another like this one that found different behavior for StarData and AstronomicalData:

http://community.wolfram.com/groups/-/m/t/885567

The only result there was to use parallelisation; because each kernel than creates a connection and downloads data...