The cluster scalar uses total memory instead of available memory.

[ejacox]

The cluster scaler calculation to determine which machine types to start uses the total memory value. The Mesos slaves will return available memory, though, which might not be sufficient to run the jobs. This leads to the situation where the clusterScalar will not spin up additional machines because it thinks that there are enough, but some jobs will hang forever since the memory offered by Mesos is insufficient. This occurred in issue #2078, which was resolved by lowering memory requirements, but did not fix the underlying issue.

┆Issue is synchronized with this Jira Story
┆friendlyId: TOIL-233