Record-blocking is a technique that significantly reduces the number of network flows.
It does this by returning a block of rows from the server on the first FETCH request for a cursor. Subsequent FETCH requests are retrieved from the local block of rows, rather then going to the server each time. This technique dramatically increases performance when it is properly used. The default settings should be sufficient for most situations.
A change to one of the record-blocking parameters can make a significant difference when the performance of your environment is approaching the exponential threshold that is illustrated in Introduction to client/server performance. For example, assume that an environment has n decision-support clients doing some amount of work with large queries, typically returning 1 MB of data.
At the opposite extreme is a scenario where users consistently ask for large amounts of data, but typically never examine more than a few rows. The overhead of returning 32KB of rows when only a few are needed could degrade performance. Setting the BLOCKSIZE or BlockSizeKB connection string keyword to a lower value, setting the BLOCKFETCH connection string keyword to 0 (Use ODBC blocking) or disabling record blocking altogether, might actually increase performance.
It is important to note that, as always in client/server, performance results may vary. You might make changes to these parameters and not realize any difference. This may indicate that your performance bottleneck is not the client request queue at the server. This parameter gives you one more tool to use when your users start objecting.