Thank you very much, This is exactly the answer I was searching for!
- Thodoris > On 22 May 2018, at 21:10, Joseph Wu <[email protected]> wrote: > > 1) DRF is based on the _current_ allocation of resources (from the master's > perspective) rather than a historical allocation of resources. > > 2) So when a new cluster is started, all frameworks will have a current > allocation of 0. And assuming all else (like quotas, roles, and weights) are > equivalent (or not set to anything), then your 2 frameworks would receive > roughly equal shares of offers. > > 3) As of right now, there is no way for the framework to directly influence > the number of offers received in a single call. The best approach to getting > offers on multiple machines is to hold onto the offers (i.e. not accepting > nor declining them) until your necessary conditions have been met. > >> On Tue, May 22, 2018 at 2:56 AM, Thodoris Zois <[email protected]> wrote: >> Hello list, >> >> I have some questions about resource offers for Mesos and I am >> experiencing some problems that I hope somebody will be able to help. >> >> 1) The allocation module of Mesos master uses DRF (according to >> previous allocation history) and decides which framework will get an >> offer, and how many resources will be offered. Is this right? >> >> 2) Assume that a Mesos cluster starts for the very first time and 2 >> frameworks join. None of the frameworks has a job to submit, they just >> wait and get offers. What is the policy to send resource offers since >> master does not know anything about previous allocations? >> >> 3) I got a Mesos cluster with 5 machines and 1 framework only. Is there >> any way to force Mesos send everytime all the five resource offers to >> my framework? I have seen that when the framework registers for the >> first time, it gets a list of offers that include all 5 machines. >> However, if it does not accept them the next round of offers contains >> only 2 or even 1 machine, depending on the time that framework has >> declined. >> >> Thank you very much for your response, >> Any help is appreciated! >> >> - Thodoris >
