Allow to benchmark multi-machine clusters

Currently, it is only possible to benchmark on a single machine (except for the special case where we run with --pipeline=benchmark-only which (a) puts the burden of provisioning on the user and (b) does not gather system metrics (like CPU usage or index size). Rally should be able to run benchmarks also on clusters across multiple machines.

In this ticket we want to collect the high-level ideas around making this possible. Work should be done in smaller, more focused tickets. At least these areas need to change:

• Provisioning (consider also #42 and #117)
• Gathering metrics (we need to gather them on the machines that are being benchmarked)
• Reporting: We currently report system level metrics only once (e.g. CPU usage) and need to reconsider that (although we will prepare for that case already in #54)

Note: Distribution of the load generator is handled separately in #257.

[danielmitterdorfer]

We’ll implement this gradually and implement support for single-machine clusters in #184 first. Then we can extend Rally to multi-machine clusters. This will reduce the risk of adding too many changes at once (as we then also need an ability to define cluster topology).

[mxplusb]

Gotcha, thanks!