Hi, I tried to answer your questions inline below: Overall I would recommend you to have a look at the following paper explaining some of the background of scheduling and architecture : http://mesos.berkeley.edu/mesos_tech_report.pdf and https://www.cs.berkeley.edu/~alig/papers/drf.pdf.
Regards, Joerg On Sat, Mar 14, 2015 at 2:15 PM, Kenneth Hui <[email protected]> wrote: > All, > > I am new to the Mesos project and have been reading up on the technology > and have questions for which I have not been able to find answers. I would > appreciate any answers to the questions below or links to where I can find > the answers myself. > > 1. How does the Master determine the order in which it offers > resources? Is it a simple round-robin, taking into account filters, or is > there a algorithm that is followed? > > This is determined by the allocator where the standard is the Dominant Resource Fairness explained in the second paper above. For more details see here http://mesos.apache.org/documentation/latest/allocation-module/ > > 1. Are there any mechanisms where a framework can initiate an offer > request so that it receives the next resource offer or do frameworks always > have to wait for the Master to decide the order of resource offers? > > Framework (Scheduler) have to wait for the Master to offer them ressources. See links in above answer for more details. > > 1. When a task is executed across a cluster, does the Master always > try to distribute the task across all slave nodes with sufficient available > resources or would it ever run all task on a single node in the cluster if > it has sufficient resources? > > The nodes announces free resources which the master then offers to the registered framework. So it depends on how many offers the framework(s) accept. > > 1. How does Mesos respond if all its slave nodes run out of physical > resources and are unable to offer sufficient resources for an extended > period of time? > > As long as the slave (process) is reachable from the master this is ok as the node is participating in the cluster (with 'old' jobs). > > Thank you for your help. > > > -- > Ken > > >
