multiple sendmsg and single user-kernel switch

Hi all,
I want to use the sendto (or sendmsg) for sending raw packets on a raw
socket directly to the NIC (passing the kernel protocol stack).
In addition to that i need so send a lot of packets (100-200Kpps) so i
have to do that fast.
Currently the kernel-user switching is killing me and i want to batch
some packets before calling send() - to reduce the user-kernel switch
overhead.
Passing a batched buffer to the sendto() cause a probelm - the NIC have
to get one message at a time because it has to calculate the L2 CRC for
each packet (or msg). so batched buffer is not acceptable.

I would like to get some help please. any one have any idea how to
reduce the user-kernel overhead (or sending a batched buffer so the
kernel will call sendto for each buffer)?

Thanks,
Aviv