On 5/28/10, Lynn W. Taylor <[email protected]> wrote: > > > On 5/28/2010 2:59 PM, Rattledagger wrote: > >> Don't point to the wrong problem, user manually resetting project very >> many >> times is an unlikely scenario. While this does in some way generate >> "ghost-wu", >> the most common reason for "ghost-wu" is due to connection-problem. If >> scheduling-server gets a clients request for work, and scheduling-server >> sends >> one or more wu's to user, but internet-connection drops-out before client >> gets >> the reply, you've got a "ghost-wu". And let's face it, internet-connection >> dropping-out is by no means an unfrequent occorrence, and in most >> instances >> users has no control of this, so it's not something they're doing >> "wrong"... > > I haven't looked at the BOINC protocol, but.... > > The server should not consider the work unit "assigned" until the remote > end acks (accepts) the assignment. > > That keeps the server from thinking something is assigned when the > client doesn't have it. > > Three-steps probably isn't out of the question. TCP uses a three step > connection sequence to prevent one side from thinking that the > connection is open while the other side is still waiting.
Currently: Client asks for N seconds of work, server gives some tasks, and there is potentially no more communication between them until the client finishes the tasks and reports them. Martin and I had already suggested a protocol where the client asks for work, the server "offers" tasks, and the client then has to say if it wants them or not. For example, the client could ask for work to all projects it's attached to, and based on what is available, decide which ones to accept. See "[boinc_dev] Work Scheduling (pt 2, Cooperative Scheduling?)", from April 29, 2009. -- Nicolas _______________________________________________ boinc_dev mailing list [email protected] http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev To unsubscribe, visit the above URL and (near bottom of page) enter your email address.
