On 31.01.14 10:30, Rytis Slatkevičius wrote:
We had pretty much the same problem. If the cluster nodes retain their network card MAC addresses, version 7.2.33 or newer generates host cpid from
MAC address and therefore the scheduler should always match the "new" host to the old record in the database.
Thanks for the note.
However, I'm afraid this doesn't work for us: Other clusters run one client per "slot" (more or less continuously), i.e. CPU core. Therefore we
enabled <multiple_clients_per_host> in the project config, which (as far as I understand the code) prevents the scheduler from searching for
matching hosts when attaching a new client.
I am also not sure how the VMs of the cluster and their MAC addresses are managed. In case the number of possible MAC adresses is limited, it might be
worth to add some custom code to determine the hostid from the MAC address (or CPID in case of recent clients) for specific user accounts (possibly
only for single-CPU hosts).
I'm still not sure, though, what happens when one task is shifted to a different VM, then the VM that first downloaded it starts another client (and
attaches another host). There would be two clients running with the same hostid but different sequence numbers, which would likely lead to a problem
e.g. when reporting. One might tackle this special case by adding additional, even more special custom code that only re-uses the id of hosts that
have no tasks in progress.
Best,
Bernd
_______________________________________________
boinc_dev mailing list
boinc_dev@ssl.berkeley.edu
http://lists.ssl.berkeley.edu/mailman/listinfo/boinc_dev
To unsubscribe, visit the above URL and
(near bottom of page) enter your email address.
