[Beowulf] Q: IB message rate & large core counts (per node)?
atchley at myri.com
Wed Feb 24 18:49:38 PST 2010
On Feb 23, 2010, at 6:16 PM, Brice Goglin wrote:
> Greg Lindahl wrote:
>>> now that I'm inventorying ignorance, I don't really understand why RDMA
>>> always seems to be presented as a big hardware issue. wouldn't it be
>>> pretty easy to define an eth or IP-level protocol to do remote puts,
>>> gets, even test-and-set or reduce primitives, where the interrupt handler
>>> could twiddle registered blobs of user memory on the target side?
>> That approach is called Active Messages, and can be bolted on to
>> pretty much every messaging implementation. Doesn't OpenMX provide
>> that kind of interface?
> Open-MX offers what MX offers: no explicit RDMA interface, only 2-sided.
> But something similar to a remote get is used internally for large
> messages. It wouldn't be hard to mplement some RDMA-like features in
> such a software-only model like Mark said above.
Don't forget the unexpected handler which can provide some Active Message behavior.
More information about the Beowulf