Some questions of DFL to support clustering
yanyh at cs.uh.edu
Mon Nov 15 14:16:01 PST 2004
Per my understanding, dragonfly designed to be high performance for SMP, and
later scale up to a single system image for clustering. In current work, is
there any topics covering effecient inter-node (inter-box) communication? I
mean not the fast TCP/IP, but some more effecient and less complex protocol
for fast message passing between two OSes. And also some topics for
collective communication in clustering or SMP?
Also, seem to me that DFL put the scheduler to userland which I like very
much because lots of application-specific co-scheduling protocol in
clustering can be developed on it. But how much the scheduling is push to
userland, completely, or partially.
User-land process/thread control: any new interface from userland to
control a thread/process? These should be important if put scheduler to
userland. Can I put a process/thread purposely to a waiting state or
Thanks very much.
More information about the Kernel