9.2. Throughput of Non-Posted Reads
To support a high throughput for read data, you must analyze the overall delay from the time the Application Layer issues the read request until all of the completion data is returned. The Application Layer must be able to issue enough read requests, and the read completer must be capable of processing these read requests quickly enough (or at least offering enough non-posted header credits) to cover this delay.
However, much of the delay encountered in this loop is well outside the IP core and is very difficult to estimate. PCI Express switches can be inserted in this loop, which makes determining a bound on the delay more difficult.
Nevertheless, maintaining maximum throughput of completion data packets is important. Endpoints must offer an infinite number of completion credits. Endpoints must buffer this data in the RX buffer until the Application Layer can process it. Because the Endpoint is no longer managing the RX buffer for Completions through the flow control mechanism, the Application Layer must manage the RX buffer by the rate at which it issues read requests.
To determine the appropriate settings for the amount of space to reserve for completions in the RX buffer, you must make an assumption about the length of time until read completions are returned. This assumption can be estimated in terms of an additional delay, beyond the FC Update Loop Delay, as discussed in the section Throughput of Posted Writes. The paths for the read requests and the completions are not exactly the same as those for the posted writes and FC Updates in the PCI Express logic. However, the delay differences are probably small compared with the inaccuracy in the estimate of the external read to completion delays.
With multiple completions, the number of available credits for completion headers must be larger than the completion data space divided by the maximum packet size. Instead, the credit space for headers must be the completion data space (in bytes) divided by 64, because this is the smallest possible read completion boundary. Setting the RX Buffer space allocation—Desired performance for received completions to High under the System Settings heading when specifying parameter settings configures the RX buffer with enough space to meet this requirement. You can adjust this setting up or down from the High setting to tailor the RX buffer size to your delays and required performance.