github.com/smintz/nomad@v0.8.3/website/source/docs/internals/architecture.html.md (about) 1 --- 2 layout: "docs" 3 page_title: "Architecture" 4 sidebar_current: "docs-internals-architecture" 5 description: |- 6 Learn about the internal architecture of Nomad. 7 --- 8 9 # Architecture 10 11 Nomad is a complex system that has many different pieces. To help both users and developers of Nomad 12 build a mental model of how it works, this page documents the system architecture. 13 14 ~> **Advanced Topic!** This page covers technical details 15 of Nomad. You do not need to understand these details to 16 effectively use Nomad. The details are documented here for 17 those who wish to learn about them without having to go 18 spelunking through the source code. 19 20 # Glossary 21 22 Before describing the architecture, we provide a glossary of terms to help 23 clarify what is being discussed: 24 25 * **Job** - A Job is a specification provided by users that declares a workload for 26 Nomad. A Job is a form of _desired state_; the user is expressing that the job should 27 be running, but not where it should be run. The responsibility of Nomad is to make sure 28 the _actual state_ matches the user desired state. A Job is composed of one or more 29 task groups. 30 31 * **Task Group** - A Task Group is a set of tasks that must be run together. For example, a 32 web server may require that a log shipping co-process is always running as well. A task 33 group is the unit of scheduling, meaning the entire group must run on the same client node and 34 cannot be split. 35 36 * **Driver** – A Driver represents the basic means of executing your **Tasks**. 37 Example Drivers include Docker, Qemu, Java, and static binaries. 38 39 * **Task** - A Task is the smallest unit of work in Nomad. Tasks are executed by drivers, 40 which allow Nomad to be flexible in the types of tasks it supports. Tasks 41 specify their driver, configuration for the driver, constraints, and resources required. 42 43 * **Client** - A Client of Nomad is a machine that tasks can be run on. All clients run the 44 Nomad agent. The agent is responsible for registering with the servers, watching for any 45 work to be assigned and executing tasks. The Nomad agent is a long lived process which 46 interfaces with the servers. 47 48 * **Allocation** - An Allocation is a mapping between a task group in a job and a client 49 node. A single job may have hundreds or thousands of task groups, meaning an equivalent 50 number of allocations must exist to map the work to client machines. Allocations are created 51 by the Nomad servers as part of scheduling decisions made during an evaluation. 52 53 * **Evaluation** - Evaluations are the mechanism by which Nomad makes scheduling decisions. 54 When either the _desired state_ (jobs) or _actual state_ (clients) changes, Nomad creates 55 a new evaluation to determine if any actions must be taken. An evaluation may result 56 in changes to allocations if necessary. 57 58 * **Server** - Nomad servers are the brains of the cluster. There is a cluster of servers 59 per region and they manage all jobs and clients, run evaluations, and create task allocations. 60 The servers replicate data between each other and perform leader election to ensure high 61 availability. Servers federate across regions to make Nomad globally aware. 62 63 * **Regions and Datacenters** - Nomad models infrastructure as regions and 64 datacenters. Regions may contain multiple datacenters. Servers are assigned to 65 a specific region, managing state and making scheduling decisions within that 66 region. Multiple regions can be federated together. For example, you may 67 have a `US` region with the `us-east-1` and `us-west-1` datacenters, 68 connected to the `EU` region with the `eu-fr-1` and `eu-uk-1` datacenters. 69 Requests that are made between regions are forwarded to the appropriate servers. 70 Data is _not_ replicated between regions. 71 72 * **Bin Packing** - Bin Packing is the process of filling bins with items in a way that 73 maximizes the utilization of bins. This extends to Nomad, where the clients are "bins" 74 and the items are task groups. Nomad optimizes resources by efficiently bin packing 75 tasks onto client machines. 76 77 # High-Level Overview 78 79 Looking at only a single region, at a high level Nomad looks like this: 80 81 [![Regional Architecture](/assets/images/nomad-architecture-region.png)](/assets/images/nomad-architecture-region.png) 82 83 Within each region, we have both clients and servers. Servers are responsible for 84 accepting jobs from users, managing clients, and [computing task placements](/docs/internals/scheduling.html). 85 Each region may have clients from multiple datacenters, allowing a small number of servers 86 to handle very large clusters. 87 88 In some cases, for either availability or scalability, you may need to run multiple 89 regions. Nomad supports federating multiple regions together into a single cluster. 90 At a high level, this setup looks like this: 91 92 [![Global Architecture](/assets/images/nomad-architecture-global.png)](/assets/images/nomad-architecture-global.png) 93 94 Regions are fully independent from each other, and do not share jobs, clients, or 95 state. They are loosely-coupled using a gossip protocol, which allows users to 96 submit jobs to any region or query the state of any region transparently. Requests 97 are forwarded to the appropriate server to be processed and the results returned. 98 Data is _not_ replicated between regions. 99 100 The servers in each region are all part of a single consensus group. This means 101 that they work together to elect a single leader which has extra duties. The leader 102 is responsible for processing all queries and transactions. Nomad is optimistically 103 concurrent, meaning all servers participate in making scheduling decisions in parallel. 104 The leader provides the additional coordination necessary to do this safely and 105 to ensure clients are not oversubscribed. 106 107 Each region is expected to have either three or five servers. This strikes a balance 108 between availability in the case of failure and performance, as consensus gets 109 progressively slower as more servers are added. However, there is no limit to the number 110 of clients per region. 111 112 Clients are configured to communicate with their regional servers and communicate 113 using remote procedure calls (RPC) to register themselves, send heartbeats for liveness, 114 wait for new allocations, and update the status of allocations. A client registers 115 with the servers to provide the resources available, attributes, and installed drivers. 116 Servers use this information for scheduling decisions and create allocations to assign 117 work to clients. 118 119 Users make use of the Nomad CLI or API to submit jobs to the servers. A job represents 120 a desired state and provides the set of tasks that should be run. The servers are 121 responsible for scheduling the tasks, which is done by finding an optimal placement for 122 each task such that resource utilization is maximized while satisfying all constraints 123 specified by the job. Resource utilization is maximized by bin packing, in which 124 the scheduling tries to make use of all the resources of a machine without 125 exhausting any dimension. Job constraints can be used to ensure an application is 126 running in an appropriate environment. Constraints can be technical requirements based 127 on hardware features such as architecture and availability of GPUs, or software features 128 like operating system and kernel version, or they can be business constraints like 129 ensuring PCI compliant workloads run on appropriate servers. 130 131 # Getting in Depth 132 133 This has been a brief high-level overview of the architecture of Nomad. There 134 are more details available for each of the sub-systems. The [consensus protocol](/docs/internals/consensus.html), 135 [gossip protocol](/docs/internals/gossip.html), and [scheduler design](/docs/internals/scheduling.html) 136 are all documented in more detail. 137 138 For other details, either consult the code, ask in IRC or reach out to the mailing list. 139