Sorry for delay, Konstantin Boudnik.
I attached document. It's draft and it's not been published yet. This describes top-level concept of using GGHA. Please read this first.
Some additional comments:
The both processors (MR and GGFS) are running by default. It doesn't need to configure it if you don't need to use complicated dual modes of GGFS.
More detailed info about this read here:
GridGain cluster nodes don't have master and slave roles. By default all nodes find each other in current subnetwork and each node can play master role.
I think the main issue is the classpath for Hadoop components. These jars must be added to all client application and yarn nodes:
Outside of BigTop we usual recommend to add links to this jars into <hadoop_distribution>/share/hadoop/common/lib directory.
And about memory usage.
By default GGFS uses only heap memory is allocated in Xmx java option and MR generally uses off-heap memory. To change this value, you should export JVM_OPTS with all options are defined in bin/include/service.sh and change Xmx value if it's needed.