Edit how-it-works
Project: http://git-wip-us.apache.org/repos/asf/incubator-myriad/repo Commit: http://git-wip-us.apache.org/repos/asf/incubator-myriad/commit/5ca89f36 Tree: http://git-wip-us.apache.org/repos/asf/incubator-myriad/tree/5ca89f36 Diff: http://git-wip-us.apache.org/repos/asf/incubator-myriad/diff/5ca89f36 Branch: refs/heads/master Commit: 5ca89f36805075a9ba3fb0675960fe79b5b4d33d Parents: b8c45fd Author: rharris <rhar...@maprtech.com> Authored: Fri Aug 7 16:22:23 2015 -0700 Committer: rharris <rhar...@maprtech.com> Committed: Fri Aug 7 16:22:23 2015 -0700 ---------------------------------------------------------------------- docs/how-it-works.md | 43 ++++++++++++++++++++++++++----------------- 1 file changed, 26 insertions(+), 17 deletions(-) ---------------------------------------------------------------------- http://git-wip-us.apache.org/repos/asf/incubator-myriad/blob/5ca89f36/docs/how-it-works.md ---------------------------------------------------------------------- diff --git a/docs/how-it-works.md b/docs/how-it-works.md index 096419c..3dfbfc3 100644 --- a/docs/how-it-works.md +++ b/docs/how-it-works.md @@ -1,43 +1,52 @@ -# How it works ? +# How it Works? -### Introduction +## Introduction -Myriad allows Mesos and YARN to co-exist and share resources with Mesos as the resource manager for the datacenter. Sharing resources between these two resource allocation systems improves overall cluster utilization and avoids statically partitioning resources amongst two separate clusters/resource managers. +Myriad allows Mesos and YARN to co-exist and share resources with Mesos as the resource manager for the datacenter. Sharing resources between these two resource allocation systems improves overall cluster utilization and avoids statically partitioning resources amongst two separate clusters and resource managers. -Running two resource managers independently results in a statically partitioned datacenter as shown below: +The following diagram shows two resource managers running independently which results in a statically partitioned datacenter:  -### Advertising resources: Mesos Slave and YARNâs Node Manager +## Advertising Resources: Mesos Slave and YARNâs Node Manager -Mesos Slave and YARNâs Node Manager are processes that run on the host OS, both advertises available resources to Mesos Master and YARNâs Resource Manager respectively. Each process can be configured to advertise a subset of resources. We can leverage this ability, in conjunction with cgroups, to allow Mesos Slave and YARNâs Node Manager to co-exist on a node. The diagram below showcases a node running YARN NodeManager as a Mesos Slave task: +The Mesos Slave and YARNâs Node Manager are processes that run on the host OS. Both processes advertise available resources to Mesos Master and YARNâs Resource Manager respectively. Each process can be configured to advertise a subset of resources. This ability is leveraged, in conjunction with cgroups, to allow Mesos Slave and YARNâs Node Manager to co-exist on a node. + +The following diagram showcases a node running YARN NodeManager as a Mesos Slave task:  -Let Mesos Slave be the processes that advertises all of a nodeâs resources (8 CPUs, 16 GB RAM) to Mesos Master. Now, let's start YARN Node Manager as a Mesos Task. This task is allotted (4 CPUs and 8 GB RAM), and the Node Manager is configured to only advertise 3 CPUs and 7 GB RAM. The Node Manager is also configured to mount the YARN containers under the [cgroup hierarchy](cgroups.md) which stems from a Mesos task. Ex: +* Mesos Slave processes advertises all of a nodeâs resources (8 CPUs, 16 GB RAM) to Mesos Master. +* The YARN Node Manager is started as a Mesos Task. This task is allotted (4 CPUs and 8 GB RAM) and the Node Manager is configured to only advertise 3 CPUs and 7 GB RAM. +* The Node Manager is also configured to mount the YARN containers under the [cgroup hierarchy](cgroups.md) which stems from a Mesos task. + +For example: + ```bash /sys/fs/cgroup/cpu/mesos/node-manager-task-id/container-1 ``` -Doing above allows Mesos Slave and Node Manager to co-exist on the same node, in a non-intrusive way. Our architecture leverages this strategy; we'll explore how this fits into the bigger picture. -### High Level Design +## High Level Design + +One way to avoid static partitioning and to enable resource sharing when running two resource managers, is to let one resource manager be in absolute control of the datacenterâs resources. The other resource manager then manages a subset of resources, allocated to it through the primary resource manager. -One way to avoid static partitioning and to enable resource sharing when running two resource managers, is to let one resource manager be in absolute control of datacenterâs resources. The other resource manager then manages a subset of resources, allocated to it through the primary resource manager. Let's consider a scenario where Mesos is used as the resource manager for the datacenter. In the diagram below, both, Mesos and YARN, can schedule tasks on any node. +The following diagram shows a scenario where Mesos is used as the resource manager for the datacenter which allows both Mesos and YARN to schedule tasks on any node.  -Let's look at how we can achieve above, that is, how we can run YARN along side Mesos. The diagram below gives an overview: +The following diagram gives an overview of how YARN can run along side Mesos:  -Each node in the cluster has both daemons, Mesos slave and YARN node manager, installed. By default, the Mesos slave daemon is started on each node and advertises all available resources to the Mesos Master. +Each node in the cluster has both daemons, Mesos Slave and YARN Node Manager, installed. By default, the Mesos slave daemon is started on each node and advertises all available resources to the Mesos Master. -Myriad can launch NodeManager as a task under Mesos Slave, let's look at how: +Myriad launches NodeManager as a task under Mesos Slave: 1. Myriad makes a decision to launch a new NodeManager. - a. It passes the required configuration and task launch information to Mesos Master which forwards that to the Mesos Slave(s). b. Mesos Slave launches Myriad Executor which will manage the lifecycle of the NodeManager. - c. Myriad Executor upon launch, configures Node Manager appropriately like specifying CPU and memory to advertise, cgroups hierarchy, etc. and then launches it. For ex: In the above diagram, Node Manager is allotted 2.5 CPU and 2.5 GB RAM. -2. NodeManager, upon startup, advertises configured resources to YARN's Resource Manager. In the above example, 2 CPU and 2 GB RAM are advertised. The rest of the resources are used by Myriad Executor and NodeManager process itself to run. -3. YARN's Resource Manager can launch containers now, via this Node Manager. The launched containers will be mounted under the configured cgroup hierarchy, as explained in [cgroups doc](cgroups.md). + * It passes the required configuration and task launch information to the Mesos Master which forwards that to the Mesos Slave(s). + * Mesos Slave launches Myriad Executor which manages the lifecycle of the NodeManager. + * Myriad Executor upon launch, configures Node Manager (for example, specifying CPU and memory to advertise, cgroups hierarchy, and so on) and then launches it. For example: In the previous diagram, Node Manager is allotted 2.5 CPU and 2.5 GB RAM. +2. NodeManager, upon startup, advertises configured resources to YARN's Resource Manager. In the previous example, 2 CPU and 2 GB RAM are advertised. The rest of the resources are used by the Myriad Executor and NodeManager processes to run. +3. YARN's Resource Manager can launch containers now, via this Node Manager. The launched containers are mounted under the configured cgroup hierarchy. See [cgroups doc](cgroups.md) for more information.
