Partition-based scans are more likely to create additional demands on pools, since multiple worker processes may be performing asynchronous prefetching on different allocation units. On partitioned tables on multiple devices, the per-server and per-engine I/O limits are less likely to be reached, but the per-pool limits are more likely to limit prefetching.
Once a parallel query is parsed and compiled, it launches its worker processes. If a table with 4 partitions is being scanned by 4 worker processes, each worker process attempts to prefetch all the pages in its first allocation unit. For the performance of this single query, the most desirable outcome is that the size and limits on the 16K pool are sufficiently large to allow 124 (31*4) asynchronous prefetch requests, so all of the requests succeed. Each of the worker processes scans the pages in cache quickly, moving onto new allocation units and issuing more prefetch requests for large numbers of pages.