Hello, thank you for your elaborate response.
On Mon, Aug 3, 2009 at 5:22 PM, Tzachi Dar<[email protected]> wrote: > I guess that if you can live with latency of ~5us you should be using SDP > which will give you the minimum development effort. development is only a minor issue -- unless we're talking developer-years here :-). > > If you want to be in the area of ~1us latency you should be using one of the > following 3 options IBAL ND or winverbs. I guess that once it comes to > latency this are the best choices. It appears that ND is MS-only and results in even more vendor-lock in, which we would like to avoid. From the remaining two, IBAL seems to be the more mature implementation, whereas WV has a bright future. I personally prefer the latter, all else being equal, especially when this API is closer to its linux counterpart. We're currently running XP x64, so 32bit support will not be an issue. > Please note that to get the best latency one should be using RDMA write and > pool on the memory for completion (see the ib_write_lat program for an > example). speaking of example code, is there anything in the trunk demonstrating the usage of the winverbs api? The docs mention ibv_read_lat and friends but I'm unable to find these, or anything else using the winverbs header file. What about uDAPL? It claims to be a thin layer abstracting RDMA-enabled communication from the underlying hardware and the API is OS-independent. Additionally, the endpoint state diagram from the specification is very close to what we currently have in our tcp/ip overlapped I/O implementation. Where's the drawback? Thanks again, Thomas _______________________________________________ ofw mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/ofw
