Generally, the number of packets being transferred is more important than the size of the packets. “Network” performance also includes the time needed by the CPU and operating system to process a network packet. This per-packet overhead affects performance the most. Larger packets reduce the overall overhead costs and achieve higher physical throughput, provided that you have enough data to be sent.
The following big transfer sources may benefit from large packet sizes:
Bulk copy
readtext and writetext commands
select statements with large result sets
There is always a point at which increasing the packet size will not improve performance, and may in fact decrease performance, because the packets are not always full. Although there are analytical methods for predicting that point, it is more common to vary the size experimentally and plot the results. If you conduct such experiments over a period of time and conditions, you can determine a packet size that works well for a lot of processes. However, since the packet size can be customized for every connection, specific experiments for specific processes can be beneficial.
The results can be significantly different between applications. Bulk copy might work best at one packet size, while large image data retrievals might perform better at a different packet size.
If testing shows that some specific applications can achieve better performance with larger packet sizes, but that most applications send and receive small packets, clients need to request the larger packet size.
