[Openib-windows] SRP msg size

[Jan Bottorff]

>I am looking into the SRP performance and I notice that the target use
4K
>msg size packets for RDMA transactions.
>>From the code I see that the initiator get gather/scatter list of 4K
>elements from storport 
>and send them to the target.

>Do you know why StorPort use elements with 4K size? And not 16K/32K ?
>Can we configure Storport to use larger msg size?

I've been told some OS's (Linux?) will allocate physical memory in small
clusters of virtual memory pages (which are 4k) and as a result the I/O
transfers will end up being to 16k physically contiguous regions.
Windows allocates memory on a virtual memory page size basis, so
transfer fragments will generally be to page size fragments. This is the
architecture of the OS memory manager, and is not a configurable option.
You could map 4k fragments into larger regions using FMR's or something,
although mapping isn't free, so may or may not be more expensive than 4K
RDMA transactions. 

- Jan