SPI speed at very small packet sizes (<24 bit)

SPI data transfer takes a long time if you use small packets. In my case I’m sending 24 bits in one package at 1 MHz.

The SPI configuration is set up properly, csAssertFn or csDessertFn set to NULL and deselectTime set to 0.

There’s an initial delay of about 85us until the transaction (SPI CLK) starts. Then the transaction takes 24us as expected.
Afterwards there’s another huge delay until the SPI transfer completes. I managed to reduce that delay to 35us by modifying spi_master.c (suppressing NAuWait for small SPI packets, see http://www.digi.com/support/forum/viewthread_thread,1200_offset,0)

It seems that NASpiReadWrite incorporates an excessive overhead due to DMA setup. I already talked to the DIGI support, it’s a common behaviour of that platform and SPI driver.

Did anyone of you modify the driver to improve the timing ?
Can you please give me some hints what to change to decrease the overhead ?

Best regards,
Thomas