[MESOS-6375] Support hierarchical resource allocation roles.

• Type: Epic
• Status: Reviewable
• Priority: Major
• Resolution: Unresolved
• Component/s: allocation
• Labels: mesosphere multitenancy resource-management
• Epic Name: hierarchical roles
• Target Version/s: 1.4.0
• Assignee: Neil Conway
• Reporter: Benjamin Mahler
• Created: 12/Oct/16 01:26
• Updated: 21/Jun/19 23:43

Description

Currently mesos provides a non-hierarchical resource allocation model, in which all roles are siblings of one another.

Organizations often have a need for hierarchical resource allocation constraints, whether for fair sharing of resources or for specifying quota constraints.

Consider the following fair sharing hierarchy based on "shares":

              ^                           ^
            /   \                       /   \
          /       \                   /       \
       eng (3)   sales (1)  =>   eng (75%)  sales (25%)
         ^                          ^
       /   \                      /   \
     /       \                  /       \
  ads (2)    build (1)      ads (66%)  build (33%)

The hierarchy specifies that the engineering organization should get 3x as many resources as sales, and within these resources the ads team should get 2x as many resources as the build team. The implication of this is that, if the ads team is not using some of its resources, the build team and engineering organization will be able to use these resources before the sales organization can. Without a hierarchy, the resources unused by the ads team would be re-distributed among all other roles (rather than only its siblings).

Quota can also apply in a hierarchical manner:

                    ^
                  /   \
                /       \
       eng (90 cpus)   sales (10 cpus)
             ^
           /   \
         /       \
 ads (50 cpus)   build (10 cpus)

See https://people.eecs.berkeley.edu/~alig/papers/h-drf.pdf for some discussion w.r.t. sharing resources in a hierarchical model.

Issue Links

• blocks: MESOS-7136 Eliminate fair sharing between frameworks within a role.
• blocks: MESOS-7577 Remove GPU_RESOURCES capability and master flag `--filter-gpu-resources={true|false}`
• blocks: MESOS-7579 Deprecate GPU_RESOURCES capability and master flag `--filter-gpu-resources={true|false}`
• is a child of: MESOS-7575 Support reservation refinement for hierarchical roles.
• is related to: MESOS-6432 Roles with quota assigned can "game" the system to receive excessive resources.
• is related to: MESOS-1763 Add support for frameworks to receive resources for multiple roles.
• relates to: MESOS-6774 Role sorter and quota role sorter can have more copies of share resources in allocations than in total.
• relates to: MESOS-7390 Evaluate and optimize sorter performance
• relates to: MESOS-7576 Add master flag `--filter-gpu-resources={true|false}`