...
Anchor
| Anchor | 1009849 | 1009849
|---|
| Table of Contents |
Along with the reception of UDP traffic, you should test your device driver for TCP traffic. The following are two tests that measure the driver performance under different conditions.Anchor
Test 1: Receive TCP Test with using NDIT
...
To achieve the best possible throughput, you might have to increase the number of receive descriptors and receive buffers. On the other hand, it is also possible to reserve too many buffers for reception. To find out the ideal number of descriptors and buffers, there are two things you need to measure.Anchor
First, you must determine the rate at which the target can receive data. This value, in bits per second, will be referred to as the bandwidth. You can obtain this value by running the receive UDP test.Anchor
Second, you must determine the round trip time (RTT) of a message between the test station and the target. This is achieved by sending an ICMP echo request to the target and measuring the RTT of the reply. You can use ping (or preferably fping) to acquire this value.Anchor
Then take these two value and multiply them to determine the Bandwidth-Delay Product.Anchor
BDP (bytes) = total_available_bandwidth (KBytes/sec) • round_trip_time (ms)Anchor
The BDP is approximately equal to the Receive TCP Window Size. It is recommended to round up the calculated value to a multiple of the Maximum Segment Size (MSS), typically 1460 bytes.Anchor
For example, the bandwidth of a test station-target link is 32.461 Mbps as found by the Receive UDP test. The measured RTT is 0.9 millisecond. It gives us a BDP of 28315 bits (32.461 Mbps x 0.9 ms) or 3539 bytes. Rounding up this result to a multiple of the MSS value gives us 4380 bytes. If the combined size of the receive buffers cannot hold the BDP, the receive buffers must be increased in order to have optimal performances. It is important to increase the number of receive descriptors (RxDescNbr) accordingly.anchor
...
| Panel |
|---|
...
|
...
|
There are four options for the TCP receive test:anchor
The size of the receive socket window on the target host.
...
Buffer SizeAnchor
The length of the buffer to transmit to the target.
...
Tx Window SizeAnchor
The size of the transmit socket window on the test station.
...
Test DurationAnchor
Located in the General Options tab.
...
The TCPs test parameters must be adjusted to the following: both Buffer Size and Tx Window Size should be set to 65500. These settings will minimize the overhead of socket creation on the test station, and make full use of the available processing power of the target. The Receive Window Size must be set to the value of the BDP, rounded up to a multiple of the MSS.Anchor
You should set the TCP Receive Window size in net_cfg.h as follows:Anchor
#define NET_TCP_CFG_RX_WIN_SIZE_OCTET (RxDescNbr * 1460)uAnchor
Expected Results
...
- Driver throughput should be optimized.
...
- Few transitory errors.
...
See the section on transitory errors on
...
for more information.
...
- No retransmission.
...
See the section on retransmission on
...
for more details.
...
- The messages “Window update”, “Zero window” and “Window probe” are acceptable.
...
These messages are part of a flow control mechanism that prevents the receiver from getting more packets that it can actually handle, or for the transmitter to wait indefinitely for acknowledgement to resume the transmission.
...
...
- No buffer leaks.
...
...
for more details.
...
- Logging performance results (with the target directly connected, and networked).