We've observed databases with very large buffer caches where Serial Scans don't make use of Smart Scan when that would have executed faster: improvements to the decision making for Serial Scans have been made under bug 31626438. This fix is back-portable.
A key difference between PQ and Serial is that as part of granule generation PQ sums the sizes of all the partitions that have not been pruned and passes that total size to the buffer cache decision making logic. Because the entire size to be scanned is considered, we make an accurate determination of smart scan benefits and the risk of cache thrashing.
Serial Scans on partitioned tables do not involved the coordinator and have no opportunity to get the larger picture, instead they start work immediately so each partition is considered one at a time and only that one partition's size is considered by the decision for using Buffer Cache or Direct Read (and hence offload). In the presence of very large buffer caches any given partition can fail the "Is Medium" test (or even the "Is Small" test) and so not get offloaded.
In order to avoid this situation an upper bound of 100MB for using a buffer cache scan has been implemented for any serially scanned segment that:
Any partitions larger than 100 MB will now automatically use Direct Read and hence offload on Exadata.
See also: Part 1
See also: Part 2
Another case to watch out for is when NSMTIO tracing shows HWM_NOT_FOUND and then choosing a Buffer Cache scan when a Direct Read offloaded scan would have been faster. This can happen when a PQ query gets executed serially (NB: this is NOT the downgrade to serial case, this is still PQ but on a single thread). In this case the coordinator again does not have the opportunity to process all the partitions and as part of that gather the High Water Mark (HWM) for each segment and checkpoint them so we fall back on buffer cache scans. A fix for this is currently being investigated.
I have consistently advised against mixing block sizes in a database without a compelling reason backed up by empirical evidence, but for those who must the "Is Medium Table" logic for whether to use buffer cache or direct read has been improved when the database has more than one block size in use. This is tracked by bug 24655250 and fixed in 20.1.
See also Random thoughts on block sizes