In Dask distributed documentation, they have the following information:
For example Dask developers use this ability to build in data locality
when we communicate to data-local storage systems like the Hadoop File
System. When users use high-level functions like
dask.dataframe.read_csv('hdfs:///path/to/files.*.csv') Dask talks to
the HDFS name node, finds the locations of all of the blocks of data,
and sends that information to the scheduler so that it can make
smarter decisions and improve load times for users.
However, it seems that the get_block_locations() was removed from the HDFS fs backend, so my question is: what is the current state of Dask regarding to HDFS ? Is it sending computation to nodes where data is local ? Is it optimizing the scheduler to take into account data locality on HDFS ?
Quite right, with the appearance of arrow's HDFS interface, which is now preferred over hdfs3, the consideration of block locations is no longer part of workloads accessing HDFS, since arrow's implementation doesn't include the get_block_locations() method.
However, we already wanted to remove the somewhat convoluted code which made this work, because we found that the inter-node bandwidth on test HDFS deployments was perfectly adequate that it made little practical difference in most workloads. The extra constrains on the size of the blocks versus the size of the partitions you would like in-memory created an additional layer of complexity.
By removing the specialised code, we could avoid the very special case that was being made for HDFS as opposed to external cloud storage (s3, gcs, azure) where it didn't matter which worker accessed which part of the data.
In short, yes the docs should be updated.
Related
Can someone explain using the word count example, why Spark would be faster than Map Reduce?
bafna's answer provides the memory-side of the story, but I want to add other two important facts：DAG and ecosystem
Spark uses "lazy evaluation" to form a directed acyclic graph (DAG) of consecutive computation stages. In this way, the execution plan can be optimized, e.g. to minimize shuffling data around. In contrast, this should be done manually in MapReduce by tuning each MR step. (It would be easier to understand this point if you are familiar with the execution plan optimization in RDBMS or the DAG-style execution of Apache Tez)
Spark ecosystem has established a versatile stack of components to handle SQL, ML, Streaming, Graph Mining tasks. But in the hadoop ecosystem you have to install other packages to do these individual things.
And I want to add that, even if your data is too big for main memory, you can still use spark by choosing to persist you data on disk. Although by doing this you give up the advantages of in-memory processing, you can still benefit from the DAG execution optimization.
Some informative answers on Quora:
here and here.
I think there are three primary reasons.
The main two reasons stem from the fact that, usually, one does not run a single MapReduce job, but rather a set of jobs in sequence.
One of the main limitations of MapReduce is that it persists the full dataset to HDFS after running each job. This is very expensive, because it incurs both three times (for replication) the size of the dataset in disk I/O and a similar amount of network I/O. Spark takes a more holistic view of a pipeline of operations. When the output of an operation needs to be fed into another operation, Spark passes the data directly without writing to persistent storage. This is an innovation over MapReduce that came from Microsoft's Dryad paper, and is not original to Spark.
The main innovation of Spark was to introduce an in-memory caching abstraction. This makes Spark ideal for workloads where multiple operations access the same input data. Users can instruct Spark to cache input data sets in memory, so they don't need to be read from disk for each operation.
What about Spark jobs that would boil down to a single MapReduce job? In many cases also these run faster on Spark than on MapReduce. The primary advantage Spark has here is that it can launch tasks much faster. MapReduce starts a new JVM for each task, which can take seconds with loading JARs, JITing, parsing configuration XML, etc. Spark keeps an executor JVM running on each node, so launching a task is simply a matter of making an RPC to it and passing a Runnable to a thread pool, which takes in the single digits of milliseconds.
Lastly, a common misconception probably worth mentioning is that Spark somehow runs entirely in memory while MapReduce does not. This is simply not the case. Spark's shuffle implementation works very similarly to MapReduce's: each record is serialized and written out to disk on the map side and then fetched and deserialized on the reduce side.
I wrote a script that analyzes a lot of files on an AWS cluster.
Running it on the cloud seems to be slower than I expected - the filesystem is shared via NFS, so the round-trip through the network seems to be the limiting step here. Bottom line - the processing power of the cluster is limited by the speed of the internal network which is considerably slower than the speed of the SSD the data is located in.
How would you optimize the cluster so that IO intensive jobs will run efficiently?
There isn't much you can do given the circumstances.
Obviously the speed of the NFS itself is the drawback.
Consider:
Chunking - grab only pieces of the files required to do as much as possible
Copying locally - create a locking mechanism and copy the file in full locally, process, push back. This can require a lot of work (what if the worker gives up and doesnt clear the lock)
Optimize the NFS share - increase IO throughput by clustering the NFS, raiding it, etc.
With a remote FS you want to limit the amount of back and forth. You can be creative, but creative can be a problem in itself.
I intend to setup spark cluster on EC2. How much resources spark master instance actually needs? Since master is not involved in processing any of the tasks can it be the smallest EC2 instance?
This obviously depends on what kinds of jobs you're planning to run, how big is the cluster etc, so in that sense the advice to simply try different configurations is good. However, in my purely personal experience the driver instance should be at least at the level of the slave instances. This is mainly due to two reasons.
First of all, there are times when you need the result of the job in a single place. Maybe you just don't want to spend time combining files, maybe you need the results in some specific order which would be hard to achieve in a distributed way etc. but this means the driver should be able to hold all the data (as rdd.collect gathers the results to the driver instance).
Second of all, many of the shuffle-based operations seem to require a lot of memory from the driver. I'm not exactly sure about the details of why this happens (if anyone knows, please do share) but I can't count the number of times I've seen reduceyKey causing an out of memory error from the driver.
Edit: I have assumed you were using Spark's spark-ec2 script, which I believe does install the NameNode in the master instance. If the NameNode is not installed at the master intance, however, my answer has no validity as correctly pointed by #DemetriKots in the comments.
Although the master instance is not involved in data processing, it plays a major role during the management of the workload and resource allocation, e.g (all info is taken from the sources):
NameNode
The NameNode is the centerpiece of an HDFS file system. It keeps the directory tree of all files in the file system, and tracks where across the cluster the file data is kept. It does not store the data of these files itself.
Client applications talk to the NameNode whenever they wish to locate a file, or when they want to add/copy/move/delete a file. The NameNode responds the successful requests by returning a list of relevant DataNode servers where the data lives.
This (look for Hardware Recommendations for Hadoop on the left index) Hortonworks document specifies some recommendations for the master instance in a Hadoop cluster. While it might not be adequate for the slave instances (due to Spark's memory usage), I would say it can be useful in the case of the master instance in a Spark cluster.
I am a new learner of Hadoop.
While reading about Apache HDFS I learned that HDFS is write once file system. Some other distributions ( Cloudera) provides append feature. It will be good to know rational behind this design decision. In my humble opinion, this design creates lots of limitations on Hadoop and make it suitable for limited set of problems( problems similar to log analytic).
Experts comment will help me to understand HDFS in better manner.
HDFS follows the write-once, read-many approach for its files and applications. It assumes that a file in HDFS once written will not be modified, though it can be access ‘n’ number of times (though future versions of Hadoop may support this feature too)! At present, in HDFS strictly has one writer at any time. This assumption enables high throughput data access and also simplifies data coherency issues. A web crawler or a MapReduce application is best suited for HDFS.
As HDFS works on the principle of ‘Write Once, Read Many‘, the feature of streaming data access is extremely important in HDFS. As HDFS is designed more for batch processing rather than interactive use by users. The emphasis is on high throughput of data access rather than low latency of data access. HDFS focuses not so much on storing the data but how to retrieve it at the fastest possible speed, especially while analyzing logs. In HDFS, reading the complete data is more important than the time taken to fetch a single record from the data. HDFS overlooks a few POSIX requirements in order to implement streaming data access.
http://www.edureka.co/blog/introduction-to-apache-hadoop-hdfs/
There are three major reasons that HDFS has the design it has,
HDFS was designed by slavishly copying the design of Google's GFS, which was intended to support batch computations only
HDFS was not originally intended for anything but batch computation
Design a real distributed file system that can support high performance batch operations as well as real-time file modifications is difficult and was beyond the budget and experience level of the original implementors of HDFS.
There is no inherent reason that Hadoop couldn't have been built as a fully read/write file system. MapR FS is proof of that. But implementing such a thing was far outside of the scope and capabilities of the original Hadoop project and the architectural decisions in the original design of HDFS essentially preclude changing this limitation. A key factor is the presence of the NameNode since HDFS requires that all meta-data operations such as file creation, deletion or file length extensions round-trip through the NameNode. MapR FS avoids this by completely eliminating the NameNode and distributing meta-data throughout the cluster.
Over time, not having a real mutable file system has become more and more annoying as the workload for Hadoop-related systems such as Spark and Flink have moved more and more toward operational, near real-time or real-time operation. The responses to this problem have included
MapR FS. As mentioned ... MapR implemented a fully functional high performance re-implementation of HDFS that includes POSIX functionality as well as noSQL table and streaming API's. This system has been in performance for years at some of the largest big data systems around.
Kudu. Cloudera essentially gave up on implementing viable mutation on top of HDFS and has announced Kudu with no timeline for general availability. Kudu implements table-like structures rather than fully general mutable files.
Apache Nifi and the commercial version HDF. Hortonworks also has largely given up on HDFS and announced their strategy as forking applications into batch (supported by HDFS) and streaming (supported by HDF) silos.
Isilon. EMC implemented the HDFS wire protocol as part of their Isilon product line. This allows Hadoop clusters to have two storage silos, one for large-scale, high-performance, cost-effective batch based on HDFS and one for medium-scale mutable file access via Isilon.
other. There are a number of essentially defunct efforts to remedy the write-once nature of HDFS. These include KFS (Kosmix File System) and others. None of these have significant adoption.
An advantage of this technique is that you don't have to bother with synchronization. Since you write once, your reader are guaranteed that the data will not be manipulated while they read.
Though this design decision does impose restrictions, HDFS was built keeping in mind efficient streaming data access.
Quoting from Hadoop - The Definitive Guide:
HDFS is built around the idea that the most efficient data processing pattern is a
write-once, read-many-times pattern. A dataset is typically generated or copied
from source, and then various analyses are performed on that dataset over time.
Each analysis will involve a large proportion, if not all, of the dataset, so the time
to read the whole dataset is more important than the latency in reading the first
record.
Using HDFS for Nuodb as storage. Would this have a performance impact?
If I understand correctly, HDFS is better suited for batch mode or write once and read many times, types of application. Would it not increase the latency for record to be fetch in case it needs to read from storage?
On top of that HDFS block size concept, keep the file size small that would increase the network traffic while data is being fetch. Am I missing something here? Please point out the same.
How would Nuodb manage these kind of latency gotchas?
Good afternoon,
My name is Elisabete and I am the Technical Support Engineer over at NuoDB. I believe that I may have just answered this via your post on our own forum, but I'm responding here as well for anyone else who's curious.
First... a mini lesson on NuoDB architecture/layout:
The most basic NuoDB set-up includes:
Broker Agent
Transaction Engine (TE)
Storage Manager (SM) connected to an Archive Directory
Broker Agents keep track of all the moving parts in the domain (collection of machines hosting NuoDB processes) and provide client applications with connection information for the next available Transaction Engine.
Transaction Engines process incoming SQL requests and manage transactions.
Storage Managers read and write data to and from "disk" (Archive Directory)
All of these components can reside on a single machine, but an optimal set up would have them spread across multiple host machines (allowing each process to take full advantage of the host's available CPU/RAM). Also, while it's possible to run with just one of each component, this is a case where more is definitely more. Additional Brokers provide resiliency, additional TE's increase performance/speed and additional SM's ensure durability.
Ok, so now lets talk about Storage:
This is the "Archive Directory" that your storage manager is writing to. Currently, we support three modes of storage:
Local Files System
Amazon Web Services: Simple Storage volume (S3), Elastic Block Storage (EBS)
Hadoop Distributed Files System (HDFS)
So, to elaborate on how NuoDB works with HDFS... it doesn't know about the multiple machines that the HDFS layer is writing to. As far as the SM is concerned, it is reading and writing data atoms to a single directory. The HDFS layer decides how to then distribute and retrieve data to and from the cluster of machines it resides over.
And now to finally address the question of latency:
Here's the thing, whenever we introduce a remote storage device, we inevitably introduce some amount of additional latency because the SM now has further to go when reading/writing atoms to/from memory. HDFS likely adds a bit more, because now it needs to do it's magic divvying up, distributing, retrieving and reassembling data. Add to that discrepancy in network speed, etc.
I imagine that the gained disk space outweighs the cost in travel time, but this is something you'd have to decide on a case by case basis.
Now, all of that said... I haven't mentioned that TE and SM's both have the ability to cache data to local memory. The size of this cache is something you can set, when starting up each process. NuoDB uses a combination of Multi-Version Concurrency Control (MVCC) and a near constant stream of communication between all of the processes, to ensure that data held in cache is kept up to date with all of the changes happening within the system. Garbage Collection also kicks in and clears out atoms in a Least Recently Used order, when the cache grows close to hitting its limit.
All of this helps reduce latency, because the TE's can hold onto the data they reference most often and grab copies of data they don't have from sibling TE's. When they do resort to asking the SM's for data, there's a chance that the SM (or one of its sibling SM's) has a copy of the requested data in local cache, saving itself the trip out to the Archive Directory.
Whew.. that was a lot and I absolutely glossed over more than a few concepts. These topics are covered in greater depth via the new suite of white papers (and the new "green book") available on our main website. I'm currently also working on some visual guides, to help explain all of this.
If you'd like to know more about NuoDB or if I didn't quite answer your question.... please reach out to me directly via the NuoDB Community Forums (I respond to posts there, a bit faster).
Thank you,
Elisabete
Technical Support Engineer at NuoDB