Hi,

bottelnecks are multiple copying of data and context switches. Copying 
memory is not very fast with QNX6.

A better approach is to "abuse" the resource manager just to open up the 
PCI interface in order to get access to device memory at user space.
You could request the physical addresses (base registers) when you do an 
initial "open" of the device ... then map the device memory into user space.
So you have direct access to device memory without context switches 
(message passing) and multiple copies of the same data.

An other alternative could be to use shared memory for the exchange of 
data and just to transfer references to that memory areas by very small 
read and write messages.

Best Regards

--Armin

http://www.steinhoff-automation.com


Iain Martin wrote:
> Hi, I've written a QNX PCI driver for a board that contains three duplex links so the driver has to service 6 high 
volume data streams concurrently. I implemented the driver using a Resource Manager and everything is great except for 
the performance. Is it sensible to use a Resource Manager for a high data volume PCI driver?
> When testing our device with calls to read() of 1000 bytes, performance can be ~75% of max but this drops to 20% of 
max with read() calls of 100 bytes. I am convinced that the Resource Manager is the culprit because I can test the 
device without using the Resource Manager and achieve performance of 99% of maximum with read calls 100-10000 bytes.
> Our Resource Manager read() function calls our driver read directly  so the only difference between the two tests is 
the Resource Manager messaging system.
>
> Any help or advice woudl be greatly appreciated.
>
>
>
> _______________________________________________
>
> General
> http://community.qnx.com/sf/go/post88478
>
>