On Wed, Sep 3, 2008 at 11:49 AM, AJ Guillon <[EMAIL PROTECTED]> wrote: > Hrrrm. That's really too bad because I would like to use RDMA to steal work > from other nodes along with dependent memory. If I'm loading memory for a > task on one node, and another node steals the task, the node from which the > task was stolen should stop fetching memory required for the now stolen > task. A more complex scheduler might be able to deal with this but maybe not > optimally. > > Suggestions for workarounds? > If the target device allows you to have multiple RDMA Reads in flight, you could just break the entire Read down into a series of N smaller reads, and have a fraction of that (say N/3 or N/4) in flight at a time.
When you want to "cancel" the read, simply stop issuing the remaining Reads. If your read is small that the RTT time determines its duration rather than the length read then you didn't need to cancel anyway. _______________________________________________ general mailing list [email protected] http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
