DragonFly talk at the upcoming BayLisa (15 December 2005)

Matthew Dillon dillon at apollo.backplane.com
Sun Dec 11 10:53:30 PST 2005

:On Sat, Dec 10, 2005 at 02:38:30PM -0800, Matthew Dillon wrote:
:>     Hello everyone!  I will be giving a DragonFly talk at the next
:>     Bay Lisa.  The primary focus of my talk will be a physical 
:>     characterization (latencies, overheads, etc) of MP mechanisms
:>     and algorithms implemented by DragonFly.
:>     The Bay Lisa in question will be held on December 15's 2005 7:30 p.m.
:>     to 9:30 p.m.  on Apple Campus in Cupertino (California, USA).  Site 
:>     information and directions below:
:Good, me and Jeffery Hsu can just beat your ass there then while telling
:you how much Linux rulez and stuff. :)

    If the performance I am seeing out of this Athlon 64 X2 (dual core) is
    a sign of things to come, IPIQ messaging overheads will no longer be
    an issue.  The overhead for sending an IPI is less then 400ns (its hard
    to get an exact number because the KTR logging itself adds 40-200ns
    in overhead).  The transit time from the originating cpu all the way
    through to executing the IPIQ message's function on the target cpu is
    less then 1000ns (1uS).  Even better, the pipelining case for IPIQ
    messaging, when several messages are sent at once, has an overhead
    of only 100ns per message on the sending side and 40ns per message
    on the receiving side.

    I've also got the TCP and UDP stacks instrumented, and the EM device
    as well.  Those numbers should also prove to be interesting.

					Matthew Dillon 
					<dillon at xxxxxxxxxxxxx>

More information about the Kernel mailing list