Yarn

Yarn

• 1.
  Apache Hadoop YARN:Yet Another Resource Negotiator Vinod Kumar Vavilapallih Arun C Murthyh Chris Douglasm Sharad Agarwali Mahadev Konarh Robert Evansy Thomas Gravesy Jason Lowey Hitesh Shahh Siddharth Sethh Bikas Sahah Carlo Curinom Owen O’Malleyh Sanjay Radiah Benjamin Reedf Eric Baldeschwielerh h: hortonworks.com, m: microsoft.com, i: inmobi.com, y: yahoo-inc.com, f : facebook.com Abstract The initial design of Apache Hadoop [1] was tightly fo- cused on running massive, MapReduce jobs to process a web crawl. For increasingly diverse companies, Hadoop has become the data and computational agor´a —the de facto place where data and computational resources are shared and accessed. This broad adoption and ubiquitous usage has stretched the initial design well beyond its in- tended target, exposing two key shortcomings: 1) tight coupling of a specific programming model with the re- source management infrastructure, forcing developers to abuse the MapReduce programming model, and 2) cen- tralized handling of jobs’ control flow, which resulted in endless scalability concerns for the scheduler. In this paper, we summarize the design, development, and current state of deployment of the next genera- tion of Hadoop’s compute platform: YARN. The new architecture we introduced decouples the programming model from the resource management infrastructure, and delegates many scheduling functions (e.g., task fault- tolerance) to per-application components. We provide experimental evidence demonstrating the improvements we made, confirm improved efficiency by reporting the experience of running YARN on production environ- ments (including 100% of Yahoo! grids), and confirm the flexibility claims by discussing the porting of several Copyright c 2013 by the Association for Computing Machinery, Inc. (ACM). Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the au- thor(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected]. SoCC’13, 1–3 Oct. 2013, Santa Clara, California, USA. ACM 978-1-4503-2428-1. http://dx.doi.org/10.1145/2523616.2523633 programming frameworks onto YARN viz. Dryad, Gi- raph, Hoya, Hadoop MapReduce, REEF, Spark, Storm, Tez. 1 Introduction Apache Hadoop began as one of many open-source im- plementations of MapReduce [12], focused on tackling the unprecedented scale required to index web crawls. Its execution architecture was tuned for this use case, focus- ing on strong fault tolerance for massive, data-intensive computations. In many large web companies and star- tups, Hadoop clusters are the common place where op- erational data are stored and processed. More importantly, it became the place within an or- ganization where engineers and researchers have instan- taneous and almost unrestricted access to vast amounts of computational resources and troves of company data. This is both a cause of Hadoop’s success and also its biggest curse, as the public of developers extended the MapReduce programming model beyond the capabili- ties of the cluster management substrate. A common pattern submits “map-only” jobs to spawn arbitrary pro- cesses in the cluster. Examples of (ab)uses include fork- ing web servers and gang-scheduled computation of it- erative workloads. Developers, in order to leverage the physical resources, often resorted to clever workarounds to sidestep the limits of the MapReduce API. These limitations and misuses motivated an entire class of papers using Hadoop as a baseline for unrelated environments. While many papers exposed substantial issues with the Hadoop architecture or implementation, some simply denounced (more or less ingeniously) some of the side-effects of these misuses. The limitations of the original Hadoop architecture are, by now, well un- derstood by both the academic and open-source commu- nities. In this paper, we present a community-driven effort to
• 2.
  move Hadoop pastits original incarnation. We present the next generation of Hadoop compute platform known as YARN, which departs from its familiar, monolithic architecture. By separating resource management func- tions from the programming model, YARN delegates many scheduling-related functions to per-job compo- nents. In this new context, MapReduce is just one of the applications running on top of YARN. This separa- tion provides a great deal of flexibility in the choice of programming framework. Examples of alternative pro- gramming models that are becoming available on YARN are: Dryad [18], Giraph, Hoya, REEF [10], Spark [32], Storm [4] and Tez [2]. Programming frameworks run- ning on YARN coordinate intra-application communi- cation, execution flow, and dynamic optimizations as they see fit, unlocking dramatic performance improve- ments. We describe YARN’s inception, design, open- source development, and deployment from our perspec- tive as early architects and implementors. 2 History and rationale In this section, we provide the historical account of how YARN’s requirements emerged from practical needs. The reader not interested in the requirements’ origin is invited to skim over this section (the requirements are highlighted for convenience), and proceed to Section 3 where we provide a terse description of the YARN’s ar- chitecture. Yahoo! adopted Apache Hadoop in 2006 to replace the infrastructure driving its WebMap application [11], the technology that builds a graph of the known web to power its search engine. At the time, the web graph con- tained more than 100 billion nodes and 1 trillion edges. The previous infrastructure, named “Dreadnaught,” [25] had reached the limits of its scalability at 800 machines and a significant shift in its architecture was required to match the pace of the web. Dreadnought already ex- ecuted distributed applications that resembled MapRe- duce [12] programs, so by adopting a more scalable MapReduce framework, significant parts of the search pipeline could be migrated easily. This highlights the first requirement that will survive throughout early ver- sions of Hadoop, all the way to YARN—[R1:] Scalabil- ity. In addition to extremely large-scale pipelines for Ya- hoo! Search, scientists optimizing advertising analytics, spam filtering, and content optimization drove many of its early requirements. As the Apache Hadoop com- munity scaled the platform for ever-larger MapReduce jobs, requirements around [R2:] Multi-tenancy started to take shape. The engineering priorities and intermedi- ate stages of the compute platform are best understood in this context. YARN’s architecture addresses many long- standing requirements, based on experience evolving the MapReduce platform. In the rest of the paper, we will assume general understanding of classic Hadoop archi- tecture, a brief summary of which is provided in Ap- pendix A. 2.1 The era of ad-hoc clusters Some of Hadoop’s earliest users would bring up a clus- ter on a handful of nodes, load their data into the Ha- doop Distributed File System (HDFS)[27], obtain the re- sult they were interested in by writing MapReduce jobs, then tear it down [15]. As Hadoop’s fault tolerance im- proved, persistent HDFS clusters became the norm. At Yahoo!, operators would load “interesting” datasets into a shared cluster, attracting scientists interested in deriv- ing insights from them. While large-scale computation was still a primary driver of development, HDFS also acquired a permission model, quotas, and other features to improve its multi-tenant operation. To address some of its multi-tenancy issues, Yahoo! developed and deployed Hadoop on Demand (HoD), which used Torque[7] and Maui[20] to allocate Hadoop clusters on a shared pool of hardware. Users would sub- mit their job with a description of an appropriately sized compute cluster to Torque, which would enqueue the job until enough nodes become available. Onces nodes be- come available, Torque would start HoD’s ’leader’ pro- cess on the head node, which would then interact with Torque/Maui to start HoD’s slave processes that subse- quently spawn a JobTracker and TaskTrackers for that user which then accept a sequence of jobs. When the user released the cluster, the system would automatically collect the user’s logs and return the nodes to the shared pool. Because HoD sets up a new cluster for every job, users could run (slightly) older versions of Hadoop while developers could test new features easily. Since Hadoop released a major revision every three months, 1. The flex- ibility of HoD was critical to maintaining that cadence— we refer to this decoupling of upgrade dependencies as [R3:] Serviceability. While HoD could also deploy HDFS clusters, most users deployed the compute nodes across a shared HDFS instance. As HDFS scaled, more compute clusters could be allocated on top of it, creating a virtuous cycle of increased user density over more datasets, leading to new insights. Since most Hadoop clusters were smaller than the largest HoD jobs at Yahoo!, the JobTracker was rarely a bottleneck. HoD proved itself as a versatile platform, anticipat- ing some qualities of Mesos[17], which would extend 1Between 0.1 and 0.12, Hadoop released a major version every month. It maintained a three month cycle from 0.12 through 0.19
• 3.
  the framework-master modelto support dynamic re- source allocation among concurrent, diverse program- ming models. HoD can also be considered a private- cloud precursor of EC2 Elastic MapReduce, and Azure HDInsight offerings—without any of the isolation and security aspects. 2.2 Hadoop on Demand shortcomings Yahoo! ultimately retired HoD in favor of shared MapReduce clusters due to middling resource utiliza- tion. During the map phase, the JobTracker makes ev- ery effort to place tasks close to their input data in HDFS, ideally on a node storing a replica of that data. Because Torque allocates nodes without accounting for locality,2 the subset of nodes granted to a user’s Job- Tracker would likely only contain a handful of relevant replicas. Given the large number of small jobs, most reads were from remote hosts. Efforts to combat these artifacts achieved mixed results; while spreading Task- Trackers across racks made intra-rack reads of shared datasets more likely, the shuffle of records between map and reduce tasks would necessarily cross racks, and sub- sequent jobs in the DAG would have fewer opportuni- ties to account for skew in their ancestors. This aspect of [R4:] Locality awareness is a key requirement for YARN. High-level frameworks like Pig[24] and Hive[30] of- ten compose a workflow of MapReduce jobs in a DAG, each filtering, aggregating, and projecting data at every stage in the computation. Because clusters were not re- sized between jobs when using HoD, much of the ca- pacity in a cluster lay fallow while subsequent, slimmer stages completed. In an extreme but a very common sce- nario, a single reduce task running on one node could prevent a cluster from being reclaimed. Some jobs held hundreds of nodes idle in this state. Finally, job latency was dominated by the time spent allocating the cluster. Users could rely on few heuristics when estimating how many nodes their jobs required, and would often ask for whatever multiple of 10 matched their intuition. Cluster allocation latency was so high, users would often share long-awaited clusters with col- leagues, holding on to nodes for longer than anticipated, raising latencies still further. While users were fond of many features in HoD, the economics of cluster utiliza- tion forced Yahoo! to pack its users into shared clus- ters. [R5:] High Cluster Utilization is a top priority for YARN. 2Efforts to modify torque to be “locality-aware” mitigated this ef- fect somewhat, but the proportion of remote reads was still much higher than what a shared cluster could achieve. 2.3 Shared clusters Ultimately, HoD had too little information to make intel- ligent decisions about its allocations, its resource granu- larity was too coarse, and its API forced users to provide misleading constraints to the resource layer. However, moving to shared clusters was non-trivial. While HDFS had scaled gradually over years, the Job- Tracker had been insulated from those forces by HoD. When that guard was removed, MapReduce clusters sud- denly became significantly larger, job throughput in- creased dramatically, and many of the features inno- cently added to the JobTracker became sources of criti- cal bugs. Still worse, instead of losing a single workflow, a JobTracker failure caused an outage that would lose all the running jobs in a cluster and require users to manu- ally recover their workflows. Downtime created a backlog in processing pipelines that, when restarted, would put significant strain on the JobTracker. Restarts often involved manually killing users’ jobs until the cluster recovered. Due to the com- plex state stored for each job, an implementation pre- serving jobs across restarts was never completely de- bugged. Operating a large, multi-tenant Hadoop cluster is hard. While fault tolerance is a core design principle, the surface exposed to user applications is vast. Given vari- ous availability issues exposed by the single point of fail- ure, it is critical to continuously monitor workloads in the cluster for dysfunctional jobs. More subtly, because the JobTracker needs to allocate tracking structures for every job it initializes, its admission control logic in- cludes safeguards to protect its own availability; it may delay allocating fallow cluster resources to jobs because the overhead of tracking them could overwhelm the Job- Tracker process. All these concerns may be grouped un- der the need for [R6:] Reliability/Availability. As Hadoop managed more tenants, diverse use cases, and raw data, its requirements for isolation became more stringent, but the authorization model lacked strong, scalable authentication—a critical feature for multi- tenant clusters. This was added and backported to mul- tiple versions. [R7:] Secure and auditable operation must be preserved in YARN. Developers gradually hard- ened the system to accommodate diverse needs for re- sources, which were at odds with the slot-oriented view of resources. While MapReduce supports a wide range of use cases, it is not the ideal model for all large-scale computa- tion. For example, many machine learning programs re- quire multiple iterations over a dataset to converge to a result. If one composes this flow as a sequence of MapReduce jobs, the scheduling overhead will signifi- cantly delay the result [32]. Similarly, many graph al-
• 4.
  gorithms are betterexpressed using a bulk-synchronous parallel model (BSP) using message passing to com- municate between vertices, rather than the heavy, all- to-all communication barrier in a fault-tolerant, large- scale MapReduce job [22]. This mismatch became an impediment to users’ productivity, but the MapReduce- centric resource model in Hadoop admitted no compet- ing application model. Hadoop’s wide deployment in- side Yahoo! and the gravity of its data pipelines made these tensions irreconcilable. Undeterred, users would write “MapReduce” programs that would spawn alter- native frameworks. To the scheduler they appeared as map-only jobs with radically different resource curves, thwarting the assumptions built into to the platform and causing poor utilization, potential deadlocks, and insta- bility. YARN must declare a truce with its users, and pro- vide explicit [R8:] Support for Programming Model Diversity. Beyond their mismatch with emerging framework re- quirements, typed slots also harm utilization. While the separation between map and reduce capacity prevents deadlocks, it can also bottleneck resources. In Hadoop, the overlap between the two stages is configured by the user for each submitted job; starting reduce tasks later increases cluster throughput, while starting them early in a job’s execution reduces its latency.3 The number of map and reduce slots are fixed by the cluster operator, so fallow map capacity can’t be used to spawn reduce tasks and vice versa.4 Because the two task types com- plete at different rates, no configuration will be perfectly balanced; when either slot type becomes saturated, the JobTracker may be required to apply backpressure to job initialization, creating a classic pipeline bubble. Fungi- ble resources complicate scheduling, but they also em- power the allocator to pack the cluster more tightly. This highlights the need for a [R9:] Flexible Resource Model. While the move to shared clusters improved utiliza- tion and locality compared to HoD, it also brought con- cerns for serviceability and availability into sharp re- lief. Deploying a new version of Apache Hadoop in a shared cluster was a carefully choreographed, and a re- grettably common event. To fix a bug in the MapReduce implementation, operators would necessarily schedule downtime, shut down the cluster, deploy the new bits, validate the upgrade, then admit new jobs. By conflat- ing the platform responsible for arbitrating resource us- age with the framework expressing that program, one is forced to evolve them simultaneously; when opera- tors improve the allocation efficiency of the platform, 3This oversimplifies significantly, particularly in clusters of unreli- able nodes, but it is generally true. 4Some users even optimized their jobs to favor either map or reduce tasks based on shifting demand in the cluster [28]. Node ManagerNode Manager Node Manager ResourceManager Scheduler AMService MR AM Container Container Container MPI AM .. . Container RM -- NodeManager RM -- AM Umbilical client client Client -- RM Figure 1: YARN Architecture (in blue the system components, and in yellow and pink two applications running.) users must necessarily incorporate framework changes. Thus, upgrading a cluster requires users to halt, vali- date, and restore their pipelines for orthogonal changes. While updates typically required no more than re- compilation, users’ assumptions about internal frame- work details—or developers’ assumptions about users’ programs—occasionally created blocking incompatibil- ities on pipelines running on the grid. Building on lessons learned by evolving Apache Ha- doop MapReduce, YARN was designed to address re- quirements (R1-R9). However, the massive install base of MapReduce applications, the ecosystem of related projects, well-worn deployment practice, and a tight schedule would not tolerate a radical redesign. To avoid the trap of a “second system syndrome” [6], the new ar- chitecture reused as much code from the existing frame- work as possible, behaved in familiar patterns, and left many speculative features on the drawing board. This lead to the final requirement for the YARN redesign: [R10:] Backward compatibility. In the remainder of this paper, we provide a descrip- tion of YARN’s architecture (Sec. 3), we report about real-world adoption of YARN (Sec. 4), provide experi- mental evidence validating some of the key architectural choices (Sec. 5) and conclude by comparing YARN with some related work (Sec. 6). 3 Architecture To address the requirements we discussed in Section 2, YARN lifts some functions into a platform layer respon- sible for resource management, leaving coordination of logical execution plans to a host of framework imple- mentations. Specifically, a per-cluster ResourceManager (RM) tracks resource usage and node liveness, enforces allocation invariants, and arbitrates contention among tenants. By separating these duties in the JobTracker’s charter, the central allocator can use an abstract descrip- tion of tenants’ requirements, but remain ignorant of the
• 5.
  semantics of eachallocation. That responsibility is dele- gated to an ApplicationMaster (AM), which coordinates the logical plan of a single job by requesting resources from the RM, generating a physical plan from the re- sources it receives, and coordinating the execution of that plan around faults. 3.1 Overview The RM runs as a daemon on a dedicated machine, and acts as the central authority arbitrating resources among various competing applications in the cluster. Given this central and global view of the cluster resources, it can enforce rich, familiar properties such as fairness [R10], capacity [R10], and locality [R4] across tenants. De- pending on the application demand, scheduling priori- ties, and resource availability, the RM dynamically allo- cates leases– called containers– to applications to run on particular nodes.5 The container is a logical bundle of re- sources (e.g., 2GB RAM, 1 CPU ) bound to a particular node [R4,R9]. In order to enforce and track such assign- ments, the RM interacts with a special system daemon running on each node called the NodeManager (NM). Communications between RM and NMs are heartbeat- based for scalability. NMs are responsible for monitor- ing resource availability, reporting faults, and container lifecycle management (e.g., starting, killing). The RM assembles its global view from these snapshots of NM state. Jobs are submitted to the RM via a public submis- sion protocol and go through an admission control phase during which security credentials are validated and vari- ous operational and administrative checks are performed [R7]. Accepted jobs are passed to the scheduler to be run. Once the scheduler has enough resources, the appli- cation is moved from accepted to running state. Aside from internal bookkeeping, this involves allocating a container for the AM and spawning it on a node in the cluster. A record of accepted applications is written to persistent storage and recovered in case of RM restart or failure. The ApplicationMaster is the “head” of a job, manag- ing all lifecycle aspects including dynamically increas- ing and decreasing resources consumption, managing the flow of execution (e.g., running reducers against the output of maps), handling faults and computation skew, and performing other local optimizations. In fact, the AM can run arbitrary user code, and can be writ- ten in any programming language since all communi- cation with the RM and NM is encoded using extensi- ble communication protocols6—as an example consider 5We will refer to “containers” as the logical lease on resources and the actual process spawned on the node interchangeably. 6See: https://code.google.com/p/protobuf/ the Dryad port we discuss in Section 4.2. YARN makes few assumptions about the AM, although in practice we expect most jobs will use a higher level programming framework (e.g., MapReduce, Dryad, Tez, REEF, etc.). By delegating all these functions to AMs, YARN’s ar- chitecture gains a great deal of scalability [R1], pro- gramming model flexibility [R8], and improved upgrad- ing/testing [R3] (since multiple versions of the same framework can coexist). Typically, an AM will need to harness the resources (cpus, RAM, disks etc.) available on multiple nodes to complete a job. To obtain containers, AM issues resource requests to the RM. The form of these requests includes specification of locality preferences and properties of the containers. The RM will attempt to satisfy the resource requests coming from each application according to availability and scheduling policies. When a resource is allocated on behalf of an AM, the RM generates a lease for the resource, which is pulled by a subsequent AM heartbeat. A token-based security mechanism guaran- tees its authenticity when the AM presents the container lease to the NM [R4]. Once the ApplicationMaster dis- covers that a container is available for its use, it encodes an application-specific launch request with the lease. In MapReduce, the code running in the container is either a map task or a reduce task.7 If needed, running contain- ers may communicate directly with the AM through an application-specific protocol to report status and liveness and receive framework-specific commands– YARN nei- ther facilitates nor enforces this communication. Over- all, a YARN deployment provides a basic, yet robust infrastructure for lifecycle management and monitoring of containers, while application-specific semantics are managed by each framework [R3,R8]. This concludes the architectural overview of YARN. In the remainder of this section, we provide details for each of the main components. 3.2 Resource Manager (RM) The ResourceManager exposes two public interfaces to- wards: 1) clients submitting applications, and 2) Ap- plicationMaster(s) dynamically negotiating access to re- sources, and one internal interface towards NodeMan- agers for cluster monitoring and resource access man- agement. In the following, we focus on the public pro- tocol between RM and AM, as this best represents the important frontier between the YARN platform and the various applications/frameworks running on it. The RM matches a global model of the cluster state against the digest of resource requirements reported by 7In fact, the same code is spawned by TaskTrackers in Hadoop 1.x [R10]. Once started, the process obtains its payload from the AM across the network, rather than from the local daemon.
• 6.
  running applications. Thismakes it possible to tightly enforce global scheduling properties (different sched- ulers in YARN focus on different global properties, such as capacity or fairness), but it requires the sched- uler to obtain an accurate understanding of applications’ resource requirements. Communication messages and scheduler state must be compact and efficient for the RM to scale against application demand and the size of the cluster [R1]. The way resource requests are cap- tured strikes a balance between accuracy in capturing resource needs and compactness. Fortunately, the sched- uler only handles an overall resource profile for each ap- plication, ignoring local optimizations and internal ap- plication flow. YARN completely departs from the static partitioning of resources for maps and reduces; it treats the cluster resources as a (discretized) continuum [R9]— as we will see in Section 4.1, this delivered significant improvements in cluster utilization. ApplicationMasters codify their need for resources in terms of one or more ResourceRequests, each of which tracks: 1. number of containers (e.g., 200 containers), 2. resources8 per container 2GB RAM, 1 CPU , 3. locality preferences, and 4. priority of requests within the application ResourceRequests are designed to allow users to cap- ture the full detail of their needs and/or a roll-up version of it (e.g., one can specify node-level, rack-level, and global locality preferences [R4]). This allows more uni- form requests to be represented compactly as aggregates. Furthermore, this design would allow us to impose size limits on ResourceRequests, thus leveraging the roll-up nature of ResourceRequests as a lossy compression of the application preferences. This makes communication and storage of such requests efficient for the scheduler, and it allows applications to express their needs clearly [R1,R5,R9]. The roll-up nature of these requests also guides the scheduler in case perfect locality cannot be achieved, towards good alternatives (e.g., rack-local al- location, if the desired node is busy). This resource model serves current applications well in homogeneous environments, but we expect it to evolve over time as the ecosystem matures and new re- quirements emerge. Recent and developing extensions include: explicit tracking of gang-scheduling needs, and soft/hard constraints to express arbitrary co-location or disjoint placement.9 The scheduler tracks, updates, and satisfies these re- quests with available resources, as advertised on NM heartbeats. In response to AM requests, the RM gener- 8The resource vector is designed to be extensible. 9While the RM uses locality as a weight for placing containers, network bandwidth is not explicitly modeled and reserved as in Okto- pus [5] or Orchestra [9]. ates containers together with tokens that grant access to resources. The RM forwards the exit status of finished containers, as reported by the NMs, to the responsible AMs. AMs are also notified when a new NM joins the cluster so that they can start requesting resources on the new nodes. A recent extension of the protocol allows the RM to symmetrically request resources back from an applica- tion. This typically happens when cluster resources be- come scarce and the scheduler decides to revoke some of the resources that were given to an application to maintain scheduling invariants. We use structures sim- ilar to ResourceRequests to capture the locality prefer- ences (which could be strict or negotiable). AMs have some flexibility when fulfilling such ’preemption’ re- quests, e.g., by yielding containers that are less crucial for its work (for e.g. tasks that made only little progress till now), by checkpointing the state of a task, or by migrating the computation to other running containers. Overall, this allows applications to preserve work, in contrast to platforms that forcefully kill containers to satisfy resource constraints. If the application is non- collaborative, the RM can, after waiting for a certain amount of time, obtain the needed resources by instruct- ing the NMs to forcibly terminate containers. Given the prenominate requirements from section 2, it is important to point out what the ResourceMana- ger is not responsible for. As discussed, it is not re- sponsible for coordinating application execution or task fault-tolerance, but neither is is charged with 1) pro- viding status or metrics for running applications (now part of the ApplicationMaster), nor 2) serving frame- work specific reports of completed jobs (now delegated to a per-framework daemon).10 This is consistent with the view that the ResourceManager should only handle live resource scheduling, and helps central components in YARN scale beyond the Hadoop 1.0 JobTracker. 3.3 Application Master (AM) An application may be a static set of processes, a logical description of work, or even a long-running service. The ApplicationMaster is the process that coordinates the ap- plication’s execution in the cluster, but it itself is run in the cluster just like any other container. A component of the RM negotiates for the container to spawn this boot- strap process. The AM periodically heartbeats to the RM to affirm its liveness and to update the record of its demand. After building a model of its requirements, the AM encodes its preferences and constraints in a heartbeat message to 10YARN does provide generic information about completed apps, containers etc. via a common daemon called Application History Server.
• 7.
  the RM. Inresponse to subsequent heartbeats, the AM will receive a container lease on bundles of resources bound to a particular node in the cluster. Based on the containers it receives from the RM, the AM may update its execution plan to accommodate perceived abundance or scarcity. In contrast to some resource models, the al- locations to an application are late binding: the process spawned is not bound to the request, but to the lease. The conditions that caused the AM to issue the request may not remain true when it receives its resources, but the semantics of the container are fungible and framework- specific [R3,R8,R10]. The AM will also update its re- source asks to the RM as the containers it receives affect both its present and future requirements. By way of illustration, the MapReduce AM optimizes for locality among map tasks with identical resource re- quirements. When running on HDFS, each block of in- put data is replicated on k machines. When the AM re- ceives a container, it matches it against the set of pend- ing map tasks, selecting a task with input data close to the container. If the AM decides to run a map task mi in the container, then the hosts storing replicas of mi’s input data are less desirable. The AM will update its request to diminish the weight on the other k − 1 hosts. This rela- tionship between hosts remains opaque to the RM; sim- ilarly, if mi fails, the AM is responsible for updating its demand to compensate. In the case of MapReduce, note that some services offered by the Hadoop JobTracker— such as job progress over RPC, a web interface to sta- tus, access to MapReduce-specific, historical data—are no longer part of the YARN architecture. These services are either provided by ApplicationMasters or by frame- work daemons. Since the RM does not interpret the container status, the AM determines the semantics of the success or fail- ure of the container exit status reported by NMs through the RM. Since the AM is itself a container running in a cluster of unreliable hardware, it should be resilient to failure. YARN provides some support for recovery, but because fault tolerance and application semantics are so closely intertwined, much of the burden falls on the AM. We discuss the model for fault tolerance in section 3.6. 3.4 Node Manager (NM) The NodeManager is the “worker” daemon in YARN. It authenticates container leases, manages containers’ de- pendencies, monitors their execution, and provides a set of services to containers. Operators configure it to report memory, CPU, and other resources available at this node and allocated for YARN. After registering with the RM, the NM heartbeats its status and receives instructions. All containers in YARN– including AMs– are de- scribed by a container launch context (CLC). This record includes a map of environment variables, depen- dencies stored in remotely accessible storage, security tokens, payloads for NM services, and the command necessary to create the process. After validating the au- thenticity of the lease [R7], the NM configures the envi- ronment for the container, including initializing its mon- itoring subsystem with the resource constraints speci- fied in the lease. To launch the container, the NM copies all the necessary dependencies– data files, executables, tarballs– to local storage. If required, the CLC also in- cludes credentials to authenticate the download. Depen- dencies may be shared between containers in an appli- cation, between containers launched by the same tenant, and even between tenants, as specified in the CLC. The NM eventually garbage collects dependencies not in use by running containers. The NM will also kill containers as directed by the RM or the AM. Containers may be killed when the RM reports its owning application as completed, when the scheduler decides to evict it for another tenant, or when the NM detects that the container exceeded the limits of its lease [R2,R3,R7]. AMs may request containers to be killed when the corresponding work isn’t needed any more. Whenever a container exits, the NM will clean up its working directory in local storage. When an appli- cation completes, all resources owned by its containers are discarded on all nodes, including any of its processes still running in the cluster. NM also periodically monitors the health of the phys- ical node. It monitors any issues with the local disks, and runs an admin configured script frequently that in turn can point to any hardware/software issues. When such an issue is discovered, NM changes its state to be unhealthy and reports RM about the same which then makes a scheduler specific decision of killing the con- tainers and/or stopping future allocations on this node till the health issue is addressed. In addition to the above, a NM offers local services to containers running on that node. For example, the current implementation includes a log aggregation ser- vice that will upload data written by the application to stdout and stderr to HDFS once the application completes. Finally, an administrator may configure the NM with a set of pluggable, auxiliary services. While a container’s local storage will be cleaned up after it exits, it is al- lowed to promote some output to be preserved until the application exits. In this way, a process may produce data that persist beyond the life of the container, to be managed by the node. One important use case for these services are Hadoop MapReduce applications, for which intermediate data are transferred between map and re- duce tasks using an auxiliary service. As mentioned ear- lier, the CLC allows AMs to address a payload to auxil-
• 8.
  (a) Daily jobs (b)Daily tasks Figure 2: YARN vs Hadoop 1.0 running on a 2500 nodes pro- duction grid at Yahoo!. iary services; MapReduce applications use this channel to pass tokens that authenticate reduce tasks to the shuf- fle service. 3.5 YARN framework/application writers From the preceding description of the core architecture, we extract the responsibilities of a YARN application au- thor: 1. Submitting the application by passing a CLC for the ApplicationMaster to the RM. 2. When RM starts the AM, it should register with the RM and periodically advertise its liveness and re- quirements over the heartbeat protocol 3. Once the RM allocates a container, AM can con- struct a CLC to launch the container on the correspond- ing NM. It may also monitor the status of the running container and stop it when the resource should be re- claimed. Monitoring the progress of work done inside the container is strictly the AM’s responsibility. 4. Once the AM is done with its work, it should un- register from the RM and exit cleanly. 5. Optionally, framework authors may add control flow between their own clients to report job status and expose a control plane. Even a simple AM can be fairly complex; a distributed shell example with a handful of features is over 450 lines of Java. Frameworks to ease development of YARN ap- plications exist. We explore some of these in section 4.2. Client libraries - YarnClient, NMClient, AMRMClient - ship with YARN and expose higher level APIs to avoid coding against low level protocols. An AM hardened against faults– including its own– is non-trivial. If the application exposes a service or wires a communication graph, it is also responsible for all aspects of its secure operation; YARN only secures its deployment. 3.6 Fault tolerance and availability From its inception, Hadoop was designed to run on com- modity hardware. By building fault tolerance into every layer of its stack, it hides the complexity of detection and recovery from hardware faults from users. YARN in- herits that philosophy, though responsibility is now dis- tributed between the ResourceManager and Application- Masters running in the cluster. At the time of this writing, the RM remains a single point of failure in YARN’s architecture. The RM recov- ers from its own failures by restoring its state from a per- sistent store on initialization. Once the recovery process is complete, it kills all the containers running in the clus- ter, including live ApplicationMasters. It then launches new instances of each AM. If the framework supports recovery—and most will, for routine fault tolerance— the platform will automatically restore users’ pipelines [R6]. Work is in progress to add sufficient protocol sup- port for AMs to survive RM restart. With this, AMs can continue to progress with existing containers while the RM is down, and can resync with the RM when it comes back up. Efforts are also underway to address high availability of a YARN cluster by having passive/active failover of RM to a standby node. When a NM fails, the RM detects it by timing out its heartbeat response, marks all the containers running on that node as killed, and reports the failure to all running AMs. If the fault is transient, the NM will re-synchronize with the RM, clean up its local state, and continue. In both cases, AMs are responsible for reacting to node fail- ures, potentially redoing work done by any containers running on that node during the fault. Since the AM runs in the cluster, its failure does not affect the availability of the cluster [R6,R8], but the prob- ability of an application hiccup due to AM failure is higher than in Hadoop 1.x. The RM may restart the AM if it fails, though the platform offers no support to restore the AMs state. A restarted AM synchronizing with its own running containers is also not a concern of the plat- form. For example, the Hadoop MapReduce AM will re- cover its completed tasks, but as of this writing, running tasks– and tasks that completed during AM recovery– will be killed and re-run. Finally, the failure handling of the containers them- selves is completely left to the frameworks. The RM col- lects all container exit events from the NMs and propa- gates those to the corresponding AMs in a heartbeat re- sponse. The MapReduce ApplicationMaster already lis-
• 9.
  Figure 3: YARNjobs/containers statistics on a 2500 nodes production grid at Yahoo!. tens to these notifications and retries map or reduce tasks by requesting new containers from the RM. With that, we conclude our coverage of the architec- ture and dive into YARN’s real-world installations. 4 YARN in the real-world We are aware of active use of YARN within several large companies. The following reports our experience run- ning YARN at Yahoo!. We then follow it up with a dis- cussion on few of the popular frameworks that have al- ready been ported to YARN. 4.1 YARN at Yahoo! Yahoo! upgraded its production grids from one of the stable branches of classic Hadoop to YARN. The statis- tics we report in the following are related to the last 30 days prior to upgrading for classic Hadoop, and the av- erage statistics from the upgrade time (different for each grid) up until June 13th, 2013. An important caveat in interpreting the following statistics: the results we are about to present come from a real-life upgrading experi- ence on large clusters, where the focus has been on main- taining the availability of a critical shared resource and not on scientific repeatability/consistency, as in a typical synthetic experiment. To help the reader easily interpret results, we will report a few global statistics, then fo- cus on a large grid for which hardware has (mostly) not changed before and after the upgrade, and finally char- acterize the workload shift on that specific grid. 4.1.1 YARN across all clusters In summary, after the upgrade, across all clusters, YARN is processing about 500,000 jobs daily, for a grand total of over 230 compute-years every single day. The under- lying storage exceeds 350 PB. Figure 4: Job size distribution, and resource utilization While one of the initial goals of YARN was to im- prove scalability, Yahoo! reported that they are not run- ning clusters any bigger than 4000 nodes which used be the largest cluster’s size before YARN. Significant gains in resource utilization delivered by YARN have increased the number of jobs that each grid can sustain. This has simply removed the need to scale further for the moment, and has even allowed the operational team to delay the re-provisioning of over 7000 nodes that have been decommissioned. On the other side, it appears that the log aggregation component of YARN has increased the pressure on the HDFS NameNode, especially for large jobs. The Na- meNode is now estimated to be the scalability bottleneck in Yahoo’s clusters. Fortunately, much work is ongoing in the community to both improve NameNode through- put and to limit the pressure on the NameNode by op- timizing YARN log aggregation. Scalability concerns within YARN have been observed on a few large clus- ters with massive amounts of small applications, but re- cent improvements in heartbeat handling have mitigated some of these issues. 4.1.2 Statistics on a specific cluster We will now focus on the experience of and statistics reported by Yahoo! grid operations team running YARN on one specific 2500 node grid. Figure 2 shows the before/after load the operational team is comfortable running on a 2500 machine clus- ter. This is the busiest cluster at Yahoo! which is consis- tently being pushed close to its limit. In Figure 2a, we show a significant increase in the number of jobs run- ning: from about 77k on the 1.0 version of Hadoop to roughly 100k jobs regularly on YARN. Moreover, a sus- tained throughput of 125k jobs per day was achieved on this cluster, which peaks at about 150k jobs a day (or almost about twice the number of jobs that represents the previous comfort-zone for this cluster). Average job size increased from 58 maps, 20 reducers to 85 maps, 16
• 10.
  reducers - notethat the jobs include user jobs as well as other tiny jobs that get spawned by system applications like data copy/transfer via distcp, Oozie [19] and other data management frameworks. Similarly, the number of tasks across all jobs increased from 4M on Hadoop 1.0 to about 10M on average on YARN, with a sustained load of about 12M and a peak of about 15M - see Figure 2. Another crucial metric to estimate efficiency of a clus- ter management system is the average resource utiliza- tion. Again, the shift in workload makes the statistics we are about to discuss less useful for direct comparison, but we observed a significant increase in CPU-utilization. On Hadoop 1.0, the estimated CPU utilization 11 was hovering around 320% or 3.2/16 cores pegged on each of the 2500 machines. Essentially, moving to YARN, the CPU utilization almost doubled to the equivalent of 6 continuously pegged cores per box and with peaks up to 10 fully utilized cores. In aggregate, this indicates that YARN was capable of keeping about 2.8*2500 = 7000 more cores completely busy running user code. This is consistent with the increase in number of jobs and tasks running on the cluster we discussed above. One of the most important architectural differences that par- tially explains these improvements is the removal of the static split between map and reduce slots. In Figure 3, we plot several job statistics over time: concurrently running and pending containers, jobs sub- mitted, completed, running and pending. This shows the ability of the resource manager to handle large num- ber of applications, container requests, and executions. The version of the CapacityScheduler used in this clus- ter does not use preemption yet—as this is a recently added feature. While preemption has not been tested at scale yet, we believe that careful use of preemption will significantly increase cluster utilization, we show a sim- ple microbenchmark in Section 5.3. To further characterize the nature of the workload running on this cluster, we present in Figure 4 a his- togram for different size applications depicting: 1) the total number of applications in each bucket, 2) the total amount of CPU used by applications in that bucket, and 3) the total number of containers used by applications in that bucket. As observed in the past, while a large num- ber of applications are very small, they account for a tiny fraction of cluster capacity (about 3.5 machines worth of CPU in this 2500 machines cluster). Interestingly, if we compare the slot utilization vs CPU utilization we ob- serve that large jobs seem to use more CPU for each container. This is consistent with better tuning of large jobs (e.g., consistent use of compression), and possibly longer running tasks, thus amortizing startup overhead. Overall, Yahoo! reports that the engineering effort put 11Computed by aggregating all metered CPU time and dividing by metered node uptime. in to harden YARN into a production ready, scalable system, was well worth the effort. Plans to upgrade to the latest version, 2.1-beta, continue apace. After over 36,000 years of aggregated compute-time and several months of production stress-testing, Yahoo’s operational team confirms that YARN’s architectural shift has been very beneficial for their workloads. This is well summa- rized in the following quote: “upgrading to YARN was equivalent to adding 1000 machines [to this 2500 ma- chines cluster]”. 4.2 Applications and frameworks A key requirement for YARN was to enable greater flexi- bility of programming model. This has been validated by the many programming frameworks that YARN has al- ready attracted, despite its freshly-beta status at the time of this writing. We briefly summarize some projects ei- ther native to YARN or ported to the platform to illus- trate the generality of its architecture. Apache Hadoop MapReduce already works on top of YARN with almost the same feature-set. It is tested at scale, rest of ecosystem projects like Pig, Hive, Oozie, etc. are modified to work on top of MR over YARN, together with standard benchmarks performing at par or better compared to classic Hadoop. The MapRe- duce community has made sure that applications written against 1.x can run on top of YARN in a fully binary compatible manner (mapred APIs) or just by recompil- ing (source compatibility for mapreduce APIs). Apache Tez is an Apache project (Incubator at the time of this writing) which aims to provide a generic directed- acyclic-graph (DAG) execution framework. One of its goals is to provide a collection of building blocks which can be composed into an arbitrary DAG (including a simple 2-stage (Map and Reduce) DAG to maintain compatibility with MapReduce). Tez provides query ex- ecution systems like Hive and Pig with a more natural model for their execution plan, as against forcing these plans to be transformed into MapReduce. The current focus is on speeding up complex Hive and Pig queries which typically require multiple MapReduce jobs, al- lowing to run as a single Tez job. In the future, rich fea- tures such as general support for interactive queries and generic DAGs will be considered. Spark is an open-source research project from UC Berkeley [32], that targets machine learning and inter- active querying workloads. The central idea of resilient distributed datasets (RDD) is leveraged to achieve sig- nificant performance improvements over classic MapRe- duce for this class of applications. Spark has been re- cently ported to YARN [?]. Dryad [18] provides DAG as the abstraction of execu- tion flow, and it has been integrated with LINQ [31]. The
• 11.
  version ported toYARN is 100% native C++ and C# for worker nodes, while the ApplicationMaster leverages a thin layer of Java interfacing with the ResourceMana- ger around the native Dryad graph manager. Eventually the Java layer will be substituted by direct interaction with protocol-buffer interfaces. Dryad-on-YARN is fully compatible with its non-YARN version. Giraph is a highly scalable, vertex-centric graph com- putation framework. It was originally designed to run on top of Hadoop 1.0 as a Map-only job, where one map is special and behaves as coordinator. The port to YARN of Giraph is very natural, the execution coordinator role is taken by the ApplicationMaster, and resources are re- quested dynamically. Storm is an open-source distributed, real-time pro- cessing engine, designed to scale across a cluster of ma- chines and provide parallel stream processing. A com- mon use case combines Storm for online computation and MapReduce as batch processor. By porting Storm on YARN a great deal of flexibility in resource alloca- tion can be unblocked. Moreover, the shared access to the underlying HDFS storage layer simplifies the design of multi-framework workloads. REEF meta-framework: YARN’s flexibility comes at potentially significant effort for application implemen- tors. Writing an ApplicationMaster and handling all as- pects of fault tolerance, execution flow, coordination, etc. is a non-trivial endeavor. The REEF project [10] rec- ognizes this and factors out several hard-to-build com- ponents that are common to many applications. This includes storage management, caching, fault-detection, checkpointing, push-based control flow (showcased ex- perimentally later), and container reuse. Framework de- signers can build on top of REEF and more easily than directly on YARN and reuse many common ser- vices/libraries provided by REEF. REEF design makes it suitable for both MapReduce and DAG like executions as well as iterative and interactive computations. Hoya is a Java-tool designed to leverage YARN to spin up dynamic HBase clusters[21] on demand. HBase clusters running on YARN can also grow and shrink dynamically (in our test cluster, RegionServers can be added/removed in less than 20 seconds). While the im- plications of mixing service and batch workloads in YARNare still being explored, early results from this project are encouraging. 5 Experiments In the previous section, we established the real-world success of YARN by reporting on large production de- ployments and a thriving ecosystem of frameworks. In this section, we present more specific experimental re- sults to demonstrate some of YARN’s wins. 5.1 Beating the sort record At the time of this writing, the MapReduce imple- mentation on YARN is officially 12 holding both the Daytona and Indy GraySort benchmark records, sort- ing 1.42TB/min. The same system also reported (out- side the competition) MinuteSort results sorting 1.61TB & 1.50TB in a minute, better than the current records. The experiments were run on 2100 nodes each equipped with two 2.3Ghz hexcore Xeon E5-2630, 64 GB mem- ory, and 12x3TB disks each. The summary of results is provided in the following table: Benchmark Data Type Data Size Time Rate Daytona GraySort no-skew 102.5 TB 72min 1.42TB/min Daytona GraySort skew 102.5 TB 117min 0.87TB/min Daytona MinuteSort no-skew 11497.86 GB 87.242 sec - Daytona MinuteSort skew 1497.86 GB 59.223 sec - Indy MinuteSort no-skew 1612.22 GB 58.027 sec - The full report [16] provides a detailed description of the experiments. 5.2 MapReduce benchmarks MapReduce continues to be the most important and commonly used application on top of YARN. We are pleased to report that most of the standard Hadoop benchmarks perform better on YARN in Hadoop 2.1.0 compared to the current stable Hadoop-1 release 1.2.1. The summary of the MapReduce benchmarks on a 260 node cluster comparing 1.2.1 with 2.1.0 is provided below. Each slave node is running 2.27GHz Intel(R) Xeon(R) CPU totalling to 16 cores, has 38GB physical memory, and 6x1TB 7200 RPM disks each, formatted with ext3 file-system. The network bandwidth per node is 1Gb/sec. Each node runs a DataNode and a Node- Manager with 24GB RAM allocated for containers. We run 6 maps and 3 reduces in 1.2.1, and 9 containers in 2.1.0. Each map occupies 1.5GB JVM heap and 2GB total memory, while each reduce takes 3GB heap 4GB total. JobTracker/ResourceManager run on a dedicated machine so is the HDFS NameNode. Benchmark Avg runtime (s) Throughput(GB/s) 1.2.1 2.1.0 1.2.1 2.1.0 RandomWriter 222 228 7.03 6.84 Sort 475 398 3.28 3.92 Shuffle 951 648 - - AM Scalability 1020 353/303 - - Terasort 175.7 215.7 5.69 4.64 Scan 59 65 - - Read DFSIO 50.8 58.6 - - Write DFSIO 50.82 57.74 - - Table 1: Results from canonical Hadoop benchmarks 12Source: sortbenchmark.org as of June, 2013.
• 12.
  We interpret theresults of these benchmarks as fol- lows. The sort benchmark measures the end-to-end time for a 1.52TB (6GB per node) sort in HDFS, using de- fault settings. The shuffle benchmark calibrates how fast the intermediate outputs from m maps are shuffled to n reduces using only synthetic data; records are neither read from nor written to HDFS. While the sort bench- mark would typically benefit from improvements to the HDFS data path, both benchmarks perform better on YARN primarily due to significant improvements in the MapReduce runtime itself: map-side sort improvements, a reduce client that pipelines and batches transfers of map output, and a server-side shuffle based on Netty [3]. The scan and DFSIO jobs are canonical benchmarks used to evaluate HDFS and other distributed filesys- tems run under Hadoop MapReduce; the results in ta- ble 1 are a coarse measure of the effect attributable to HDFS in our experiments. Our access to the cluster was too brief to debug and characterize the middling perfor- mance from the 2.1.0 filesystem. Despite this noise, and even though YARN’s design optimizes for multi-tenant throughput, its performance for single jobs is competi- tive with the central coordinator. The AM scalability benchmark measures single-job robustness by saturating the AM with container book- keeping duties. Table 1 includes two measurements of the MapReduce AM. The first experiment restricts avail- able resources to match the slots available to the 1.x deployment. When we remove this artificial limit and allow YARN to use the full node, its performance im- proves significantly. The two experiments estimate the overhead of typed slots. We also attribute improved per- formance to more frequent node heartbeats and faster scheduling cycles, which we discuss in greater detail be- low. Since YARN is principally responsible for distribut- ing and starting applications, we consider the scalability benchmark to be a critical metric. Some architectural choices in YARN targeted bottle- necks we observed in production clusters. As discussed in section 2.3, typed slots harm throughput by creat- ing an artificial mismatch between a fungible supply of resources on a node and the semantics of execut- ing Hadoop MapReduce tasks. While section 4.1 cov- ers the gains in aggregate workloads, we saw benefits to scheduling even single jobs. We attribute the bulk of this gain to improved heartbeat handling. In Hadoop-1, each node could heartbeat only every 30-40 seconds in large clusters due to coarse-grained locking in the Job- Tracker. Despite clever workarounds to lower latency, such as short-circuit paths for handling lightweight up- dates and adaptive delays to hasten important notifica- tions, reconciling state in the JobTracker remained the principal cause of latency in large clusters. In contrast, NodeManagers heartbeat every 1-3 seconds. The RM Figure 5: Effect of work-preserving preemption on the Capac- ityScheduler efficiency. code is more scalable, but it also solves a humbler set of constraints per request. 5.3 Benefits of preemption In Figure 5, we demonstrate a recently-added feature in YARN: the ability to enforce global properties us- ing work-preserving preemption. We ran experiments on a small (10 machine) cluster, to highlight the potential impact of work-preserving preemption. The cluster runs CapacityScheduler, configured with two queues A and B, respectively entitled to 80% and 20% of the capac- ity. A MapReduce job is submitted in the smaller queue B, and after a few minutes another MapReduce job is submitted in the larger queue A. In the graph, we show the capacity assigned to each queue under three config- urations: 1) no capacity is offered to a queue beyond its guarantee (fixed capacity) 2) queues may consume 100% of the cluster capacity, but no preemption is per- formed, and 3) queues may consume 100% of the clus- ter capacity, but containers may be preempted. Work- preserving preemption allows the scheduler to over- commit resources for queue B without worrying about starving applications in queue A. When applications in queue A request resources, the scheduler issues pre- emption requests, which are serviced by the Applica- tionMaster by checkpointing its tasks and yielding con- tainers. This allows queue A to obtain all its guaran- teed capacity (80% of cluster) in a few seconds, as op- posed to case (2) in which the capacity rebalancing takes
• 13.
  about 20 minutes.Finally, since the preemption we use is checkpoint-based and does not waste work, the job running in B can restart tasks from where they left off, and it does so efficiently. 5.4 Improvements with Apache Tez We present some rudimentary improvements when run- ning a decision support query on Apache Hive running against Apache Tez (the integration in early stages at the time of this writing). Query 12 from TPC-DS bench- mark [23], involves few joins, filters and group by aggre- gations. Even after aggressive plan level optimizations, Hive generates an execution plan consisting of multiple jobs when using MapReduce. The same query results in a linear DAG when executing against Tez, with a sin- gle Map stage followed by multiple Reduce stages. The query execution time when using MapReduce is 54 sec- onds on a 20 node cluster against 200 scale factor data, and this improves to 31 seconds when using Tez. Most of this saving can be attributed to scheduling and launch- ing overheads of multiple MapReduce jobs and avoiding the unnecessary steps of persisting outputs of the inter- mediate MapReduce jobs to HDFS. 5.5 REEF: low latency with sessions One of the key aspects of YARN is that it enables frame- works built on top of it to manage containers and com- munications as they see fit. We showcase this by lever- aging the notion of container reuse and push-based com- munications provided by REEF. The experiment is based on a simple distributed-shell application built on top of REEF. We measure the client-side latency on a com- pletely idle cluster when submitting a series of unix commands (e.g., date). The first command that is is- sued incurs the full cost of scheduling the application and acquiring containers, while subsequent commands are quickly forwarded through the Client and Appli- cationMaster to already running containers for execu- tion. The push-based messaging further reduces latency. The speedup in our experiment is very substantial, ap- proaching three orders of magnitude—from over 12sec to 31ms in average. 6 Related work Others have recognized the same limitations in the clas- sic Hadoop architecture, and have concurrently devel- oped alternative solutions, which can be closely com- pared to YARN. Among the many efforts the most closely resembling YARN are: Mesos [17], Omega [26], Corona [14], and Cosmos [8], maintained and used re- spectively by Twitter, Google, Facebook and Microsoft. These systems share a common inspiration, and the high-level goal of improving scalability, latency and pro- gramming model flexibility. The many architectural dif- ferences are a reflection of diverse design priorities, and sometimes simply the effect of different historical con- texts. While a true quantitative comparison is impossible to provide, we will try to highlight some of the archi- tectural differences and our understanding of their ratio- nale. Omega’s design leans more heavily towards dis- tributed, multi-level scheduling. This reflects a greater focus on scalability, but makes it harder to enforce global properties such as capacity/fairness/deadlines. To this goal the authors seem to rely on coordinated develop- ment of the various frameworks that will be respectful of each other at runtime. This is sensible for a closed-world like Google, but not amenable to an open platform like Hadoop where arbitrary frameworks from diverse inde- pendent sources are share the same cluster. Corona uses push based communication as opposed to the heartbeat based control-plane framework approach in YARN and other frameworks. The latency/scalability trade- offs are non-trivial and would deserve a detailed comparison. While Mesos and YARN both have schedulers at two levels, there are two very significant differences. First, Mesos is an offer-based resource manager, whereas YARN has a request-based approach. YARN allows the AM to ask for resources based on various criteria in- cluding locations, allows the requester to modify future requests based on what was given and on current us- age. Our approach was necessary to support the loca- tion based allocation. Second, instead of a per-job intra- framework scheduler, Mesos leverages a pool of central schedulers (e.g., classic Hadoop or MPI). YARN enables late binding of containers to tasks, where each individ- ual job can perform local optimizations, and seems more amenable to rolling upgrades (since each job can run on a different version of the framework). On the other side, per-job ApplicationMaster might result in greater over- head than the Mesos approach. Cosmos closely resembles Hadoop 2.0 architecturally with respect to storage and compute layers with the key difference of not having a central resource man- ager. However, it seems to be used for a single appli- cation type: Scope [8]. By virtue of a more narrow target Cosmos can leverage many optimizations such as native compression, indexed files, co-location of partitions of datasets to speed up Scope. The behavior with multiple application frameworks is not clear. Prior to these recent efforts, there is a long history of work on resource management and scheduling - Con- dor [29], Torque [7], Moab [13] and Maui [20]. Our early Hadoop clusters used some of these systems, but
• 14.
  we found thatthey could not support the MapReduce model in a first-class way. Specifically, neither the data locality nor the elastic scheduling needs of map and re- duce phases were expressible, so one was forced to al- locate “virtual” Hadoop with the attendant utilization costs discussed in section 2.1. Perhaps some of these is- sues were due to the fact that many of these distributed schedulers were originally created to support MPI style and HPC application models and running coarse-grained non-elastic workloads. These cluster schedulers do al- low clients to specify the types of processing environ- ments, but unfortunately not locality constraints which is a key concern for Hadoop. Another class of related technologies comes from the world of cloud infrastructures such as EC2, Azure, Euca- lyptus and VMWare offerings. These mostly target VM- based sharing of a cluster, and are generally designed for long running processes (as the VM boot-times overheads are prohibitive). 7 Conclusion In this paper, we summarized our recollection of the his- tory of Hadoop, and discussed how wild adoption and new types of applications has pushed the initial archi- tecture well beyond what it was designed to accomplish. We then described an evolutionary, yet profound, archi- tectural transformation that lead to YARN. Thanks to the decoupling of resource management and program- ming framework, YARN provides: 1) greater scalabil- ity, 2) higher efficiency, and 3) enables a large num- ber of different frameworks to efficiently share a cluster. These claims are substantiated both experimentally (via benchmarks), and by presenting a massive-scale produc- tion experience of Yahoo!—which is now 100% running on YARN. Finally, we tried to capture the great deal of excitement that surrounds this platform, by providing a snapshot of community activity and by briefly report- ing on the many frameworks that have been ported to YARN. We believe YARN can serve as both a solid pro- duction framework and also as an invaluable playground for the research community. 8 Acknowledgements We began our exposition of YARN by acknowledging the pedigree of its architecture. Its debt to all the individ- uals who developed, operated, tested, supported, docu- mented, evangelized, funded, and most of all, used Apa- che Hadoop over the years is incalculable. In closing, we recognize a handful. Ram Marti and Jitendranath Pan- day influenced its security design. Dick King, Krishna Ramachandran, Luke Lu, Alejandro Abdelnur, Zhijie Shen, Omkar Vinit Joshi, Jian He have made or continue to make significant contributions to the implementa- tion. Karam Singh, Ramya Sunil, Rajesh Balamohan and Srigurunath Chakravarthi have been ever helpful in test- ing and performance benchmarking. Rajive Chittajallu and Koji Noguchi helped in channeling requirements and insight from operations and user support points of view respectively. We thank Yahoo! for both support- ing YARN via massive contributions and heavy usage and also for sharing the statistics on some of its clus- ters. We thank Dennis Fetterly and Sergiy Matusevych for providing some of the experimental results on Dryad and REEF, and Mayank Bansal for helping with 2.1.0 MapReduce benchmarks. Last but not the least, Apa- che Hadoop YARN continues to be a community driven open source project and owes much of its success to the Apache Hadoop YARN and MapReduce communities— a big thanks to all the contributors and committers who have helped YARN in every way possible. A Classic Hadoop Before YARN, Hadoop MapReduce clusters were com- posed of a master called JobTracker (JT) and worker nodes running TaskTrackers (TT). Users submitted MapReduce jobs to the JT which coordinated its execu- tion across the TTs. A TT was configured by an operator with a fixed number of map slots and reduce slots. TTs periodically heartbeated into the JT to report the status of running tasks on that node and to affirm its liveness. On a heartbeat, the JT updated its state corresponding to the tasks running on that node, taking actions on behalf of that job (e.g.,scheduling failed tasks for re-execution), matching fallow slots on that node against the scheduler invariants, matching eligible jobs against available re- sources, favoring tasks with data local to that node. As the central arbiter of the compute cluster, the JT was also responsible for admission control, track- ing the liveness of TTs (to re-execute running tasks or tasks whose output becomes unavailable), launching tasks speculatively to route around slow nodes, report- ing job status to users through a web server, recording audit logs and aggregate statistics, authenticating users and many other functions; each of these limited its scal- ability. References [1] Apache hadoop. http://hadoop.apache.org. [2] Apache tez. http://incubator.apache. org/projects/tez.html. [3] Netty project. http://netty.io.
• 15.
  [4] Storm. http://storm-project.net/. [5]H. Ballani, P. Costa, T. Karagiannis, and A. I. Row- stron. Towards predictable datacenter networks. In SIGCOMM, volume 11, pages 242–253, 2011. [6] F. P. Brooks, Jr. The mythical man-month (anniver- sary ed.). Addison-Wesley Longman Publishing Co., Inc., Boston, MA, USA, 1995. [7] N. Capit, G. Da Costa, Y. Georgiou, G. Huard, C. Martin, G. Mounie, P. Neyron, and O. Richard. A batch scheduler with high level components. In Cluster Computing and the Grid, 2005. CC- Grid 2005. IEEE International Symposium on, vol- ume 2, pages 776–783 Vol. 2, 2005. [8] R. Chaiken, B. Jenkins, P.-A. Larson, B. Ramsey, D. Shakib, S. Weaver, and J. Zhou. Scope: easy and efficient parallel processing of massive data sets. Proc. VLDB Endow., 1(2):1265–1276, Aug. 2008. [9] M. Chowdhury, M. Zaharia, J. Ma, M. I. Jor- dan, and I. Stoica. Managing data transfers in computer clusters with orchestra. SIGCOMM- Computer Communication Review, 41(4):98, 2011. [10] B.-G. Chun, T. Condie, C. Curino, R. Ramakrish- nan, R. Sears, and M. Weimer. Reef: Retainable evaluator execution framework. In VLDB 2013, Demo, 2013. [11] B. F. Cooper, E. Baldeschwieler, R. Fonseca, J. J. Kistler, P. Narayan, C. Neerdaels, T. Negrin, R. Ra- makrishnan, A. Silberstein, U. Srivastava, et al. Building a cloud for Yahoo! IEEE Data Eng. Bull., 32(1):36–43, 2009. [12] J. Dean and S. Ghemawat. MapReduce: simplified data processing on large clusters. Commun. ACM, 51(1):107–113, Jan. 2008. [13] W. Emeneker, D. Jackson, J. Butikofer, and D. Stanzione. Dynamic virtual clustering with xen and moab. In G. Min, B. Martino, L. Yang, M. Guo, and G. Rnger, editors, Frontiers of High Performance Computing and Networking, ISPA 2006 Workshops, volume 4331 of Lecture Notes in Computer Science, pages 440–451. Springer Berlin Heidelberg, 2006. [14] Facebook Engineering Team. Under the Hood: Scheduling MapReduce jobs more efficiently with Corona. http://on.fb.me/TxUsYN, 2012. [15] D. Gottfrid. Self-service prorated super-computing fun. http://open. blogs.nytimes.com/2007/11/01/ self-service-prorated-super-computing-fun, 2007. [16] T. Graves. GraySort and MinuteSort at Yahoo on Hadoop 0.23. http://sortbenchmark. org/Yahoo2013Sort.pdf, 2013. [17] B. Hindman, A. Konwinski, M. Zaharia, A. Gh- odsi, A. D. Joseph, R. Katz, S. Shenker, and I. Sto- ica. Mesos: a platform for fine-grained resource sharing in the data center. In Proceedings of the 8th USENIX conference on Networked systems de- sign and implementation, NSDI’11, pages 22–22, Berkeley, CA, USA, 2011. USENIX Association. [18] M. Isard, M. Budiu, Y. Yu, A. Birrell, and D. Fet- terly. Dryad: distributed data-parallel programs from sequential building blocks. In Proceedings of the 2nd ACM SIGOPS/EuroSys European Con- ference on Computer Systems 2007, EuroSys ’07, pages 59–72, New York, NY, USA, 2007. ACM. [19] M. Islam, A. K. Huang, M. Battisha, M. Chiang, S. Srinivasan, C. Peters, A. Neumann, and A. Ab- delnur. Oozie: towards a scalable workflow man- agement system for hadoop. In Proceedings of the 1st ACM SIGMOD Workshop on Scalable Work- flow Execution Engines and Technologies, page 4. ACM, 2012. [20] D. B. Jackson, Q. Snell, and M. J. Clement. Core algorithms of the maui scheduler. In Revised Papers from the 7th International Workshop on Job Scheduling Strategies for Parallel Processing, JSSPP ’01, pages 87–102, London, UK, UK, 2001. Springer-Verlag. [21] S. Loughran, D. Das, and E. Baldeschwieler. Introducing Hoya – HBase on YARN. http://hortonworks.com/blog/ introducing-hoya-hbase-on-yarn/, 2013. [22] G. Malewicz, M. H. Austern, A. J. Bik, J. C. Dehnert, I. Horn, N. Leiser, and G. Czajkowski. Pregel: a system for large-scale graph processing. In Proceedings of the 2010 ACM SIGMOD Inter- national Conference on Management of data, SIG- MOD ’10, pages 135–146, New York, NY, USA, 2010. ACM. [23] R. O. Nambiar and M. Poess. The making of tpc- ds. In Proceedings of the 32nd international con- ference on Very large data bases, VLDB ’06, pages 1049–1058. VLDB Endowment, 2006.
• 16.
  [24] C. Olston,B. Reed, U. Srivastava, R. Kumar, and A. Tomkins. Pig Latin: a not-so-foreign language for data processing. In Proceedings of the 2008 ACM SIGMOD international conference on Man- agement of data, SIGMOD ’08, pages 1099–1110, New York, NY, USA, 2008. ACM. [25] O. O’Malley. Hadoop: The Definitive Guide, chap- ter Hadoop at Yahoo!, pages 11–12. O’Reilly Me- dia, 2012. [26] M. Schwarzkopf, A. Konwinski, M. Abd-El- Malek, and J. Wilkes. Omega: flexible, scalable schedulers for large compute clusters. In Pro- ceedings of the 8th ACM European Conference on Computer Systems, EuroSys ’13, pages 351–364, New York, NY, USA, 2013. ACM. [27] K. Shvachko, H. Kuang, S. Radia, and R. Chansler. The Hadoop Distributed File System. In Proceed- ings of the 2010 IEEE 26th Symposium on Mass Storage Systems and Technologies (MSST), MSST ’10, pages 1–10, Washington, DC, USA, 2010. IEEE Computer Society. [28] T.-W. N. Sze. The two quadrillionth bit of π is 0! http://developer.yahoo.com/blogs/hadoop/two- quadrillionth-bit-0-467.html. [29] D. Thain, T. Tannenbaum, and M. Livny. Dis- tributed computing in practice: the Condor expe- rience. Concurrency and Computation: Practice and Experience, 17(2-4):323–356, 2005. [30] A. Thusoo, J. S. Sarma, N. Jain, Z. Shao, P. Chakka, N. Z. 0002, S. Anthony, H. Liu, and R. Murthy. Hive - a petabyte scale data warehouse using Hadoop. In F. Li, M. M. Moro, S. Ghande- harizadeh, J. R. Haritsa, G. Weikum, M. J. Carey, F. Casati, E. Y. Chang, I. Manolescu, S. Mehro- tra, U. Dayal, and V. J. Tsotras, editors, Proceed- ings of the 26th International Conference on Data Engineering, ICDE 2010, March 1-6, 2010, Long Beach, California, USA, pages 996–1005. IEEE, 2010. [31] Y. Yu, M. Isard, D. Fetterly, M. Budiu, U. Erlings- son, P. K. Gunda, and J. Currey. DryadLINQ: a sys- tem for general-purpose distributed data-parallel computing using a high-level language. In Pro- ceedings of the 8th USENIX conference on Operat- ing systems design and implementation, OSDI’08, pages 1–14, Berkeley, CA, USA, 2008. USENIX Association. [32] M. Zaharia, M. Chowdhury, M. J. Franklin, S. Shenker, and I. Stoica. Spark: cluster comput- ing with working sets. In Proceedings of the 2nd USENIX conference on Hot topics in cloud com- puting, HotCloud’10, pages 10–10, Berkeley, CA, USA, 2010. USENIX Association.