There is already a JIRA ticket here https://issues.apache.org/jira/browse/MESOS-3765 tracing the fine-grained resource offers, after some discussion in community, it is suggested to reach an agreement in community before we move forward.
You can append some of your ideas there. Actually for your design document, we have discussed this in the JIRA ticket, the conclusion is that it is not good to enable framework specify the unit of allocation request as fairness cannot be guaranteed/influenced by a framework expressing its preferences, since a misbehaving/greedy framework's requests will be expressly anti-fair. As such, the requestResources API is unrelated, and we should focus on Master/operator-driven granularity adjustments. I have created a work group for this project and if you are interested in this, I can add you to that work group to move this forward. Thanks, Guangya On Thu, Dec 31, 2015 at 5:00 PM, Du, Fan <fan...@intel.com> wrote: > Hi > > Happy new year! > > Current resources offering when master performs allocation is > coarse-grained, i.e. allocating the entire unused resources on slave to one > framework on each iteration. It’s unable for framework to specify the unit > of its requested resource, and also after the offering, unneeded resources > is gave back to master for reallocation again. Firstly this process is > unnecessary as introducing latency in the transaction, moreover, with more > logical cpu cores increasing inside one physical host, such coarse-grained > could possibly bring framework on one slave only, this is not robust enough > once the slave die. Finally, with the unit information of each framework, > framework has the ability to spread workload across slaves at its will, and > choosing the right slave for allocation to avoid resource fragmentation > could become more feasible in the near future. > > Could you please review following proposal(my first proposal:) for this? > Thanks a lot. > > > https://docs.google.com/document/d/1OsdThJ758XgcPnZBcYtPIiXL23l_X1C6XswB-2yhu3k/edit?usp=sharing >
