Related Topics: Open Source Cloud

Blog Post

Realize Memory Computation in Hadoop

Memory computing will enhance Hadoop efficiency and performance

The low efficiency of Hadoop computation is an undeniable truth. We believe, one of the major reasons is that the underlying computational structure of MapReduce for Hadoop is basically of the external memory computation. The external memory computation implements the data exchange through the frequent external memory read/write. Because the efficiency of file I/O is two orders of magnitude lower than that of memory, the computational performance of Hadoop is unlikely high.

While for the normal users, they usually have a small size of cluster with only tens or scores of nodes. The cluster environment is relatively reliable and the fault probability is very low. Moreover, most realtime computations would complete quite quickly in each run. Users can always choose to recompute even on errors, not having to consider much about the fault tolerance during computation. In this case, using esProc and alike parallel computing scheme to offer double supports for both in- and external memory computations is a better choice. esProc is also based on Hadoop whose in-memory computing can be utilized for the middle-and-small scale cluster to have a much higher performance.

In the example below, we will use a typical example for grouping to illustrate how esProc implements the Hadoop memory computation. The computational goal is to summarize the sales amount on the order list by place of origin. The data are from two files on HDFS: In the sales.txt, there are a great volume of order data. The major fields are: orderID, product (product ID), amount (order value); The product.txt generates fewer data, and the main fields include proID (product ID), and origin (product origin).

The intuitive solution is like this: On the summary machine, break up sales.txt into several sections, and each section for one task. Allocate these tasks to node machine for summarizing in groups. Once the computation is done on the node machine, the result will be returned to summary machine for grouping and summarizing for a second time. The node machine is to associate sales.txt with product.txt for associative computation, and then group by origin.

esProc code is shown below:

Code1: The task decomposing and summerizing (summary machine)


Code2: Generate global variable for product table (node machine)



Code3: Associate computation and summarize by place of origion (node machine)


As can be seen, esProc coding follows the “intuitive train of thoughts” for computation. Each procedure is implemented concisely and smoothly. Most importantly, esProc has a simple structure with no hidden specifics. The actual computation is carried out step by step strictly following the code. To optimize the codes, users can modify the code in each step easily. For example: Change the granularity of task-decomposing and specify the node machine for computation.

In the following section, we will discuss the four sections of task-decomposing, data exchange, in-memory computation, and memory sharing.

Task decomposing:

As can be seen from code 1, the sales.txt can be decomposed into 40 tasks according to the computational capability of node machine, with about 1 million data for each task. With esProc, the task-decomposing scale can be customized according to their practical computing environment. For example, if the node machine is a high performance PC Server, then the 100 million pieces of data can be processsed in 10 shares, and each node for 10 million data. If the computational node is an obsolete and outdated notebook, then the data can be processed in one thousand shares, and each node can process ten thousands pieces of data. The ability to adjust the task-decomposing granularity freely will save the cost of scheduling for task-decomposing, and boost the computational performance dramatically.

By default, MapReduce decomposes the task to the minimum gratuity to address the destabilizing factor in the large scale cluster environment. Each Map task will process one record. Although the infrastructure can be modified for granularity customization, the coding is difficult and is not practical. Data decomposing by this way can address the fault at relatively less cost. However, the scheduling cost for task-decomposing is relatively great.

Data exchange:

In the code3, once the computation is done in the node machine, the result is not written into any file or sent back to the summary machine. Instead, the data exchange is done on the node machine directly. esProc is a scripting language allowing users to strike a balance between the security and the performance. For those who care more for the security on the intermediate result, esProc allows them to write the data to HDFS and then exchange; For those who care more for data exchange performance, esProc allows them to exchange the data directly.

In MapReduce, the data exchange must be done through the files to ensure the safety of intermediate result. Even if the node machine is broken down, the completed result will not disappear. In an environment of large cluster, the node machine may easily encounter such fault, which justfy this method. However, file exchanging data will definitly cause a large deal of disk IO operations and the computational performances will decline obviously.

Memory sharing:

Code2 is to read the two involved fields from the product table into the memory for computation all at once, as global variables on the node machine. Such in-memory sharing will save the time to retrieve the product table in each task. This is because every node machine will go through the computation for multiple rounds, and each round will perform the multi-threads/tasks computation. The smaller the node scale and the more computational tasks, and more obvious the performance increase will be.

MapReduce does not implement such memory sharing. That is because it is assumed that the computational node will crash frequently in the environment of large cluster, and the data obtained from the crashed memory is meaningless. In this case, it is quite safe to use the HDFS file for sharing directly. MapReduce does not support the memory sharing. Each time, users must retrieve the data from hard disk before they can use the product table. So, its efficiency is two orders of magnitude worse.

Memory computation:

As can be seen from the code 3, the product table is the global variable retrieved from the memory directly while the order table is still too great to read into memory. We use cursor to access. By this way, the efficient memory associative computation is achieved. Needless to say, if proceeding with the decomposing task, the section of order table can also be loaded into the memory. As can be seen, esProc allows for the arbitrary way to load the data. Both the file cursor method of relatively great data volume but slow spread, and the in-memory loading method for data of small volume but faster speed are enabled.

For MapReduce, the default external memory computation is to retrieve the data from file for associative computation and grouping computation. It is quite good to handle the unstable large-scale cluster environment. Although the memory buffer technique is adopted at the underlayer for MapReduce, it makes no difference on its poor performance because it is still heavily relies on the disk IO to the core. To change to the in-memory computation, users need to change the native infrastructure of MapReduce at the great development cost.

Judging from the four aspects above, we can conclude that esProc can efficiently implement the in-memory computation for Hadoop, and is suitable for users of middle and small scale cluster.

More Stories By Jessica Qiu

Jessica Qiu is the editor of Raqsoft. She provides press releases for data computation and data analytics.

Latest Stories
DX World EXPO, LLC, a Lighthouse Point, Florida-based startup trade show producer and the creator of "DXWorldEXPO® - Digital Transformation Conference & Expo" has announced its executive management team. The team is headed by Levent Selamoglu, who has been named CEO. "Now is the time for a truly global DX event, to bring together the leading minds from the technology world in a conversation about Digital Transformation," he said in making the announcement.
"Space Monkey by Vivent Smart Home is a product that is a distributed cloud-based edge storage network. Vivent Smart Home, our parent company, is a smart home provider that places a lot of hard drives across homes in North America," explained JT Olds, Director of Engineering, and Brandon Crowfeather, Product Manager, at Vivint Smart Home, in this SYS-CON.tv interview at @ThingsExpo, held Oct 31 – Nov 2, 2017, at the Santa Clara Convention Center in Santa Clara, CA.
SYS-CON Events announced today that Conference Guru has been named “Media Sponsor” of the 22nd International Cloud Expo, which will take place on June 5-7, 2018, at the Javits Center in New York, NY. A valuable conference experience generates new contacts, sales leads, potential strategic partners and potential investors; helps gather competitive intelligence and even provides inspiration for new products and services. Conference Guru works with conference organizers to pass great deals to gre...
DevOps is under attack because developers don’t want to mess with infrastructure. They will happily own their code into production, but want to use platforms instead of raw automation. That’s changing the landscape that we understand as DevOps with both architecture concepts (CloudNative) and process redefinition (SRE). Rob Hirschfeld’s recent work in Kubernetes operations has led to the conclusion that containers and related platforms have changed the way we should be thinking about DevOps and...
The Internet of Things will challenge the status quo of how IT and development organizations operate. Or will it? Certainly the fog layer of IoT requires special insights about data ontology, security and transactional integrity. But the developmental challenges are the same: People, Process and Platform. In his session at @ThingsExpo, Craig Sproule, CEO of Metavine, demonstrated how to move beyond today's coding paradigm and shared the must-have mindsets for removing complexity from the develop...
In his Opening Keynote at 21st Cloud Expo, John Considine, General Manager of IBM Cloud Infrastructure, led attendees through the exciting evolution of the cloud. He looked at this major disruption from the perspective of technology, business models, and what this means for enterprises of all sizes. John Considine is General Manager of Cloud Infrastructure Services at IBM. In that role he is responsible for leading IBM’s public cloud infrastructure including strategy, development, and offering m...
The next XaaS is CICDaaS. Why? Because CICD saves developers a huge amount of time. CD is an especially great option for projects that require multiple and frequent contributions to be integrated. But… securing CICD best practices is an emerging, essential, yet little understood practice for DevOps teams and their Cloud Service Providers. The only way to get CICD to work in a highly secure environment takes collaboration, patience and persistence. Building CICD in the cloud requires rigorous ar...
Companies are harnessing data in ways we once associated with science fiction. Analysts have access to a plethora of visualization and reporting tools, but considering the vast amount of data businesses collect and limitations of CPUs, end users are forced to design their structures and systems with limitations. Until now. As the cloud toolkit to analyze data has evolved, GPUs have stepped in to massively parallel SQL, visualization and machine learning.
"Evatronix provides design services to companies that need to integrate the IoT technology in their products but they don't necessarily have the expertise, knowledge and design team to do so," explained Adam Morawiec, VP of Business Development at Evatronix, in this SYS-CON.tv interview at @ThingsExpo, held Oct 31 – Nov 2, 2017, at the Santa Clara Convention Center in Santa Clara, CA.
To get the most out of their data, successful companies are not focusing on queries and data lakes, they are actively integrating analytics into their operations with a data-first application development approach. Real-time adjustments to improve revenues, reduce costs, or mitigate risk rely on applications that minimize latency on a variety of data sources. In his session at @BigDataExpo, Jack Norris, Senior Vice President, Data and Applications at MapR Technologies, reviewed best practices to ...
Widespread fragmentation is stalling the growth of the IIoT and making it difficult for partners to work together. The number of software platforms, apps, hardware and connectivity standards is creating paralysis among businesses that are afraid of being locked into a solution. EdgeX Foundry is unifying the community around a common IoT edge framework and an ecosystem of interoperable components.
"ZeroStack is a startup in Silicon Valley. We're solving a very interesting problem around bringing public cloud convenience with private cloud control for enterprises and mid-size companies," explained Kamesh Pemmaraju, VP of Product Management at ZeroStack, in this SYS-CON.tv interview at 21st Cloud Expo, held Oct 31 – Nov 2, 2017, at the Santa Clara Convention Center in Santa Clara, CA.
Large industrial manufacturing organizations are adopting the agile principles of cloud software companies. The industrial manufacturing development process has not scaled over time. Now that design CAD teams are geographically distributed, centralizing their work is key. With large multi-gigabyte projects, outdated tools have stifled industrial team agility, time-to-market milestones, and impacted P&L stakeholders.
"Akvelon is a software development company and we also provide consultancy services to folks who are looking to scale or accelerate their engineering roadmaps," explained Jeremiah Mothersell, Marketing Manager at Akvelon, in this SYS-CON.tv interview at 21st Cloud Expo, held Oct 31 – Nov 2, 2017, at the Santa Clara Convention Center in Santa Clara, CA.
Enterprises are adopting Kubernetes to accelerate the development and the delivery of cloud-native applications. However, sharing a Kubernetes cluster between members of the same team can be challenging. And, sharing clusters across multiple teams is even harder. Kubernetes offers several constructs to help implement segmentation and isolation. However, these primitives can be complex to understand and apply. As a result, it’s becoming common for enterprises to end up with several clusters. Thi...