Phase 1: Hadoop on Demand
The Hadoop on Demand (HOD) project was a system for provisioning and managing Hadoop MapReduce and HDFS instances on a shared cluster of commodity hardware. The Hadoop on Demand project predated and directly influenced how the developers eventually arrived at YARN’s architecture. Understanding the HOD architecture and its eventual limitations is a first step toward comprehending YARN’s motivations.
To address the multitenancy woes with the manually shared clusters from the previous incarnation (Phase 0), HOD used a traditional resource manager—Torque—together with a cluster scheduler—Maui—to allocate Hadoop clusters on a shared pool of nodes. Traditional resource managers were already being used elsewhere in high-performance computing environments to enable effective sharing of pooled cluster resources. By making use of such existing systems, HOD handed off the problem of cluster management to systems outside of Hadoop. On the allocated nodes, HOD would start MapReduce and HDFS daemons, which in turn would serve the user’s data and application requests. Thus, the basic system architecture of HOD included these layers:
- A ResourceManager (RM) together with a scheduler
- Various HOD components to interact with the RM/scheduler and manage Hadoop
- Hadoop MapReduce and HDFS daemons
- A HOD shell and Hadoop clients
A typical session of HOD involved three major steps: allocate a cluster, run Hadoop jobs on the allocated cluster, and finally deallocate the cluster. Here is a brief description of a typical HOD-user session:
Users would invoke a HOD shell and submit their needs by supplying a description of an appropriately sized compute cluster to Torque. This description included:
- The number of nodes needed
- A description of a special head-process called the RingMaster to be started by the ResourceManager
- A specification of the Hadoop deployment desired
- Torque would enqueue the request until enough nodes become available. Once the nodes were available, Torque started the head-process called RingMaster on one of the compute nodes.
- The RingMaster was a HOD component and used another ResourceManager interface to run the second HOD component, HODRing—with one HODRing being present on each of the allocated compute nodes.
- The HODRings booted up, communicated with the RingMaster to obtain Hadoop commands, and ran them accordingly. Once the Hadoop daemons were started, HODRings registered with the RingMaster, giving information about the daemons.
- The HOD client kept communicating with the RingMaster to find out the location of the JobTracker and HDFS daemons.
- Once everything was set up and the users learned the JobTracker and HDFS locations, HOD simply got out the way and allowed the user to perform his or her data crunching on the corresponding clusters.
The user released a cluster once he or she was done running the data analysis jobs.
Figure 1.1 provides an overview of the HOD architecture.
Figure 1.1 Hadoop on Demand architecture
HDFS in the HOD World
While HOD could also deploy HDFS clusters, most users chose to deploy the compute nodes across a shared HDFS instance. In a typical Hadoop cluster provisioned by HOD, cluster administrators would set up HDFS statically (without using HOD). This allowed data to be persisted in HDFS even after the HOD-provisioned clusters were deallocated. To use a statically configured HDFS, a user simply needed to point to an external HDFS instance. As HDFS scaled further, more compute clusters could be allocated through HOD, creating a cycle of increased experimentation by users over more data sets, leading to a greater return on investment. Because most user-specific MapReduce clusters were smaller than the largest HOD jobs possible, the JobTracker running for any single HOD cluster was rarely a bottleneck.
Features and Advantages of HOD
Because HOD sets up a new cluster for every job, users could run older and stable versions of Hadoop software while developers continued to test new features in isolation. Since the Hadoop community typically released a major revision every three months, the flexibility of HOD was critical to maintaining that software release schedule—we refer to this decoupling of upgrade dependencies as [Requirement 2] Serviceability.
[Requirement 2] Serviceability
The next-generation compute platform should enable evolution of cluster software to be completely decoupled from users’ applications.
In addition, HOD made it easy for administrators and users to quickly set up and use Hadoop on an existing cluster under a traditional resource management system. Beyond Yahoo!, universities and high-performance computing environments could run Hadoop on their existing clusters with ease by making use of HOD. It was also a very useful tool for Hadoop developers and testers who needed to share a physical cluster for testing their own Hadoop versions.
HOD could also be configured to upload users’ job logs and the Hadoop daemon logs to a configured HDFS location when a cluster was deallocated. The number of log files uploaded to and retained on HDFS could increase over time in an unbounded manner. To address this issue, HOD shipped with tools that helped administrators manage the log retention by removing old log files uploaded to HDFS after a specified amount of time had elapsed.
Multiple Users and Multiple Clusters per User
As long as nodes were available and organizational policies were not violated, a user could use HOD to allocate multiple MapReduce clusters simultaneously. HOD provided the list and the info operations to facilitate the management of multiple concurrent clusters. The list operation listed all the clusters allocated so far by a user, and the info operation showed information about a given cluster—Torque job ID, locations of the important daemons like the HOD RingMaster process, and the RPC addresses of the Hadoop JobTracker and NameNode daemons.
The resource management layer had some ways of limiting users from abusing cluster resources, but the user interface for exposing those limits was poor. HOD shipped with scripts that took care of this integration so that, for instance, if some user limits were violated, HOD would update a public job attribute that the user could query against.
HOD also had scripts that integrated with the resource manager to allow a user to identify the account under which the user’s Hadoop clusters ran. This was necessary because production systems on traditional resource managers used to manage accounts separately so that they could charge users for using shared compute resources.
Ultimately, each node in the cluster could belong to only one user’s Hadoop cluster at any point of time—a major limitation of HOD. As usage of HOD grew along with its success, requirements around [Requirement 3] Multitenancy started to take shape.
[Requirement 3] Multitenancy
The next-generation compute platform should support multiple tenants to coexist on the same cluster and enable fine-grained sharing of individual nodes among different tenants.
Distribution of Hadoop Software
When provisioning Hadoop, HOD could either use a preinstalled Hadoop instance on the cluster nodes or request HOD to distribute and install a Hadoop tarball as part of the provisioning operation. This was especially useful in a development environment where individual developers might have different versions of Hadoop to test on the same shared cluster.
HOD provided a very convenient mechanism to configure both the boot-up HOD software itself and the Hadoop daemons that it provisioned. It also helped manage the configuration files that it generated on the client side.
Auto-deallocation of Idle Clusters
HOD used to automatically deallocate clusters that were not running Hadoop jobs for a predefined period of time. Each HOD allocation included a monitoring facility that constantly checked for any running Hadoop jobs. If it detected no running Hadoop jobs for an extended interval, it automatically deallocated its own cluster, freeing up those nodes for future use.
Shortcomings of Hadoop on Demand
Hadoop on Demand proved itself to be a powerful and very useful platform, but Yahoo! ultimately had to retire it in favor of directly shared MapReduce clusters due to many of its shortcomings.
For any given MapReduce job, during the map phase the JobTracker makes every effort to place tasks close to their input data in HDFS—ideally on a node storing a replica of that data. Because Torque doesn’t know how blocks are distributed on HDFS, it allocates nodes without accounting for locality. The subset of nodes granted to a user’s JobTracker will likely contain only a handful of relevant replicas and, if the user is unlucky, none. Many Hadoop clusters are characterized by a small number of very big jobs and a large number of small jobs. For most of the small jobs, most reads will emanate from remote hosts because of the insufficient information available from Torque.
Efforts were undertaken to mitigate this situation but achieved mixed results. One solution was to spread TaskTrackers across racks by modifying Torque/Maui itself and making them rack-aware. Once this was done, any user’s HOD compute cluster would be allocated nodes that were spread across racks. This made intra-rack reads of shared data sets more likely, but introduced other problems. The transfer of records between map and reduce tasks as part of MapReduce’s shuffle phase would necessarily cross racks, causing a significant slowdown of users’ workloads.
While such short-term solutions were implemented, ultimately none of them proved ideal. In addition, they all pointed to the fundamental limitation of the traditional resource management software—that is, the ability to understand data locality as a first-class dimension. This aspect of [Requirement 4] Locality Awareness is a key requirement for YARN.
[Requirement 4] Locality Awareness
The next-generation compute platform should support locality awareness—moving computation to the data is a major win for many applications.
MapReduce jobs consist of multiple stages: a map stage followed by a shuffle and a reduce stage. Further, high-level frameworks like Apache Pig and Apache Hive often organize a workflow of MapReduce jobs in a directed-acyclic graph (DAG) of computations. Because clusters were not resizable between stages of a single job or between jobs when using HOD, most of the time the major share of the capacity in a cluster would be barren, waiting for the subsequent slimmer stages to be completed. In an extreme but very common scenario, a single reduce task running on one node could prevent a cluster of hundreds of nodes from being reclaimed. When all jobs in a colocation were considered, this approach could result in hundreds of nodes being idle in this state.
In addition, private MapReduce clusters for each user implied that even after a user was done with his or her workflows, a HOD cluster could potentially be idle for a while before being automatically detected and shut down.
While users were fond of many features in HOD, the economics of cluster utilization ultimately forced Yahoo! to pack its users’ jobs into shared clusters. [Requirement 5] High Cluster Utilization is a top priority for YARN.
[Requirement 5] High Cluster Utilization
The next-generation compute platform should enable high utilization of the underlying physical resources.
In a typical Hadoop workflow, MapReduce jobs have lots of maps with a much smaller number of reduces, with map tasks being short and quick and reduce tasks being I/O heavy and longer running. With HOD, users relied on few heuristics when estimating how many nodes their jobs required—typically allocating their private HOD clusters based on the required number of map tasks (which in turn depends on the input size). In the past, this was the best strategy for users because more often than not, job latency was dominated by the time spent in the queues waiting for the allocation of the cluster. This strategy, although the best option for individual users, leads to bad scenarios from the overall cluster utilization point of view. Specifically, sometimes all of the map tasks are finished (resulting in idle nodes in the cluster) while a few reduce tasks simply chug along for a long while.
Hadoop on Demand did not have the ability to grow and shrink the MapReduce clusters on demand for a variety of reasons. Most importantly, elasticity wasn’t a first-class feature in the underlying ResourceManager itself. Even beyond that, as jobs were run under a Hadoop cluster, growing a cluster on demand by starting TaskTrackers wasn’t cheap. Shrinking the cluster by shutting down nodes wasn’t straightforward, either, without potentially massive movement of existing intermediate outputs of map tasks that had already run and finished on those nodes.
Further, whenever cluster allocation latency was very high, users would often share long-awaited clusters with colleagues, holding on to nodes for longer than anticipated, and increasing latencies even further.