Heterogenous clusters
Modules can have wildly different resource needs between them. One module may need a large amount of memory, and another module might need access to GPUs. Rama provides a "label" system that allows a cluster with different kinds of hardware to be efficiently used by scheduling certain modules to certain kinds of nodes.
A supervisor can have any number of labels defined in its rama.yaml
file under the supervisor.labels
key. Here’s an example rama.yaml
using this:
zookeeper.servers:
- "1.2.3.4"
- "5.6.7.8"
conductor.host: "9.10.11.12"
local.dir: "/data/rama"
supervisor.port.range:
- 2500
- 3500
supervisor.labels:
- "gpu"
- "xlarge-cpu"
Modules can then target a particular label when being deployed. If the conductor is using the dev scheduling mode (the default), the target label is passed as part of the deploy command like this:
rama deploy \
--action launch \
--jar target/my-application.jar \
--module com.mycompany.MyModule \
--tasks 64 \
--threads 16 \
--workers 8 \
--replicationFactor 3 \
--targetLabel xlarge-cpu
The module will only be scheduled onto nodes with that label, and if not enough nodes exist with that label to satisfy the assignment the deploy will fail. --targetLabel
is optional.
The target label must also be passed on module update, similar to config overrides. This allows a label to be unset by simply not passing it. The target label does not have to be passed when scaling a module, however. This is also the same as config overrides.
If the Conductor is using the isolation scheduler, the target label is specified as part of the module’s config in the Conductor’s rama.yaml
. For example, this could be the isolation config:
conductor.assignment.mode:
type: isolation
modules:
com.mycompany.Module1:
numSupervisors: 4
targetLabel: xlarge-cpu
com.mycompany.AnotherModule: 8
com.mycompany.Module2:
numSupervisors: 12
targetLabel: gpu
com.mycompany.Module3:
numSupervisors: 10
targetLabel: xlarge-cpu
Each module must have numSupervisors
specified, but targetLabel
is optional. If the config for a module is just a number, then that is interpreted as the numSupervisors
value.
When deploying a module in isolation mode, the scheduler takes into account all the other unscheduled modules when choosing nodes. It makes a best effort to choose an assignment that also ensures the other modules will be able to be scheduled as well.