I am trying to implement data parallelism with tensorflow using the eexample replicated training documentation here
In the example code of the link above, the ps jobs seems to do nothing but blocking on the join operation:
if FLAGS.job_name == "ps":
server.join()
All the data parallelism of the computation can be taken care of by the supervisor. So what is the role of the parameter server? and can it be stopped gracefully after the jobs are done?
Related
Running a fine dask server:
Than I start pre-filling some data on the workers using publish_dataset and it looks the memory is pretty good allocated:
Next, I call futures.append(client.submit(fn, values)) in a loop and afterwards client.gather(futures) - which works! But it is not as fast as assumed, because most of the computation/memory load is beeing concentrated on 2 single processes (worker4 and worker 1):
Questions:
Why is Dask doing it like this?
Is it possible to pre-load data on specific works and than submit to the specific workers?
Should I somehow re-allocate the used memory between the workers to enable better parallelization?
Related: Dask Worker Stealing Register Worker callbacks and Dask Actors per worker
Why is Dask doing it like this?
Unfortunately I don't know enough about your problem to know why Dask is choosing to do what it's doing.
Is it possible to pre-load data on specific works and than submit to the specific workers?
Yes. See the workers= keyword to scatter and submit
Should I somehow re-allocate the used memory between the workers to enable better parallelization?
Dask should handle this for you. I recommend using scatter rather than publish_dataset, which doesn't do what you think it does (I recommend looking through the docstring for each method)
Setting the scene
I am working to make a Spark streaming application (Spark 2.2.1 with Scala) run on a Yarn cluster (Hadoop 2.7.4).
So far I managed to submit the application to the Yarn cluster with spark-submit. I can see that the receiver task starts up correctly and fetches a lot of records from the database (Couchbase Server 5.0) and I can also see that the records are divided into batches.
The question
When I look at the Streaming Statistics on the Spark Web UI, I can however see that my batches are never processed. I have seen batches with 0 records process and complete but when a batch with records start processing it never completes. One time it even got stuck on a batch with 0 records.
I even tried simplifying the output operations on the SteamingContext as much as possible. But still with the very simple output operation print() my batches are never processed. The logs does not show any warnings or errors.
Does anyone know what might be wrong? Any suggestions on how to solve this will be much appreciated.
More Info
The main class of the Spark application is built from this example (first one) from the Couchbase Spark Connector documentation combined with this example with checkpoint from the Spark Documentation.
Right now I have 3230 Active Batches (3229 queued and 1 processing) and 1 Completed Batch (that had 0 records) and the application has been running for 4 hours and 30 minutes... and another batch is added every 5 seconds.
If I look at the "thread dump" for the executors I see a lot of WAITING, TIMED WAITING and a few RUNNABLE threads. The list will fill up 3 screenshots, so i will only post it if needed.
Below you will find some screenshots from the Web UI
Executor Overview
Spark Jobs Overview
Node Overview with resources
Capacity Scheduler Overview
Per screenshot, you have 2 cores and 1 is being used for driver and another is being used for receiver. You don't have a core for the actual processing to happen. Please increase the number of cores and try again.
Refer: https://spark.apache.org/docs/latest/streaming-programming-guide.html#input-dstreams-and-receivers
If you are using an input DStream based on a receiver (e.g. sockets, Kafka, Flume, etc.), then the single thread will be used to run the receiver, leaving no thread for processing the received data. Hence, when running locally, always use “local[n]” as the master URL, where n > number of receivers to run (see Spark Properties for information on how to set the master).
I have a spark job where I need to write the output of the SQL query every micro-batch. Write is a expensive operation perf wise and is causing the batch execution time to exceed the batch interval.
I am looking for ways to improve the performance of write.
Is doing the write action in a separate thread asynchronously like shown below a good option?
Would this cause any side effects because Spark itself executes in a distributed manner?
Are there other/better ways of speeding up the write?
// Create a fixed thread pool to execute asynchronous tasks
val executorService = Executors.newFixedThreadPool(2)
dstream.foreachRDD { rdd =>
import org.apache.spark.sql._
val spark = SparkSession.builder.config(rdd.sparkContext.getConf).getOrCreate
import spark.implicits._
import spark.sql
val records = rdd.toDF("record")
records.createOrReplaceTempView("records")
val result = spark.sql("select * from records")
// Submit a asynchronous task to write
executorService.submit {
new Runnable {
override def run(): Unit = {
result.write.parquet(output)
}
}
}
}
1 - Is doing the write action in a separate thread asynchronously like shown below a good option?
No. The key to understand the issue here is to ask 'who is doing the write'. The write is done by the resources allocated for your job on the executors in a cluster. Placing the write command on an async threadpool is like adding a new office manager to an office with a fixed staff. Will two managers be able to do more work than one alone given that they have to share the same staff? Well, one reasonable answer is "only if the first manager was not giving them enough work, so there's some free capacity".
Going back to our cluster, we are dealing with a write operation that is heavy on IO. Parallelizing write jobs will lead to contention for IO resources, making each independent job longer. Initially, our job might look better than the 'single manager version', but trouble will eventually hit us.
I've made a chart that attempts to illustrate how that works. Note that the parallel jobs will take longer proportionally to the amount of time that they are concurrent in the timeline.
Once we reach that point where jobs start getting delayed, we have an unstable job that will eventually fail.
2- Would this cause any side effects because Spark itself executes in a distributed manner?
Some effects I can think of:
Probably higher cluster load and IO contention.
Jobs are queuing on the Threadpool queue instead of on the Spark Streaming Queue. We loose the ability to monitor our job through the Spark UI and monitoring API, as the delays are 'hidden' and all is fine from the Spark Streaming point of view.
3- Are there other/better ways of speeding up the write?
(ordered from cheap to expensive)
If you are appending to a parquet file, create a new file often. Appending gets expensive with time.
Increase your batch interval or use Window operations to write larger chunks of Parquet. Parquet likes large files
Tune the partition and distribution of your data => make sure that Spark can do the write in parallel
Increase cluster resources, add more nodes if necessary
Use faster storage
Is doing the write action in a separate thread asynchronously like shown below a good option?
Yes. It's certainly something to consider when optimizing expensive queries and saving their results to external data stores.
Would this cause any side effects because Spark itself executes in a distributed manner?
Don't think so. SparkContext is thread-safe and promotes this kind of query execution.
Are there other/better ways of speeding up the write?
YES! That's the key to understand when to use the other (above) options. By default, Spark applications run in FIFO scheduling mode.
Quoting Scheduling Within an Application:
By default, Spark’s scheduler runs jobs in FIFO fashion. Each job is divided into “stages” (e.g. map and reduce phases), and the first job gets priority on all available resources while its stages have tasks to launch, then the second job gets priority, etc. If the jobs at the head of the queue don’t need to use the whole cluster, later jobs can start to run right away, but if the jobs at the head of the queue are large, then later jobs may be delayed significantly.
Starting in Spark 0.8, it is also possible to configure fair sharing between jobs. Under fair sharing, Spark assigns tasks between jobs in a “round robin” fashion, so that all jobs get a roughly equal share of cluster resources. This means that short jobs submitted while a long job is running can start receiving resources right away and still get good response times, without waiting for the long job to finish. This mode is best for multi-user settings.
That means that to make a room for executing multiple writes asynchronously and in parallel you should configure your Spark application to use FAIR scheduling mode (using spark.scheduler.mode property).
You will have to configure so-called Fair Scheduler Pools to "partition" executor resources (CPU and memory) into pools that you can assign to jobs using spark.scheduler.pool property.
Quoting Fair Scheduler Pools:
Without any intervention, newly submitted jobs go into a default pool, but jobs’ pools can be set by adding the spark.scheduler.pool "local property" to the SparkContext in the thread that’s submitting them.
I'm running a hive query against a hadoop cluster of 3 nodes. And I am getting an error which says "Too many fetch failures". My hive query is:
insert overwrite table tablename1 partition(namep)
select id,name,substring(name,5,2) as namep from tablename2;
that's the query im trying to run. All i want to do is transfer data from tablename2 to tablename1. Any help is appreciated.
This can be caused by various hadoop configuration issues. Here a couple to look for in particular:
DNS issue : examine your /etc/hosts
Not enough http threads on the mapper side for the reducer
Some suggested fixes (from Cloudera troubleshooting)
set mapred.reduce.slowstart.completed.maps = 0.80
tasktracker.http.threads = 80
mapred.reduce.parallel.copies = sqrt (node count) but in any case >= 10
Here is link to troubleshooting for more details
http://www.slideshare.net/cloudera/hadoop-troubleshooting-101-kate-ting-cloudera
Update for 2020 Things have changed a lot and AWS mostly rules the roost. Here is some troubleshooting for it
https://docs.aws.amazon.com/emr/latest/ManagementGuide/emr-troubleshoot-error-resource-1.html
Too many fetch-failures
PDF
Kindle
The presence of "Too many fetch-failures" or "Error reading task output" error messages in step or task attempt logs indicates the running task is dependent on the output of another task. This often occurs when a reduce task is queued to execute and requires the output of one or more map tasks and the output is not yet available.
There are several reasons the output may not be available:
The prerequisite task is still processing. This is often a map task.
The data may be unavailable due to poor network connectivity if the data is located on a different instance.
If HDFS is used to retrieve the output, there may be an issue with HDFS.
The most common cause of this error is that the previous task is still processing. This is especially likely if the errors are occurring when the reduce tasks are first trying to run. You can check whether this is the case by reviewing the syslog log for the cluster step that is returning the error. If the syslog shows both map and reduce tasks making progress, this indicates that the reduce phase has started while there are map tasks that have not yet completed.
One thing to look for in the logs is a map progress percentage that goes to 100% and then drops back to a lower value. When the map percentage is at 100%, this does not mean that all map tasks are completed. It simply means that Hadoop is executing all the map tasks. If this value drops back below 100%, it means that a map task has failed and, depending on the configuration, Hadoop may try to reschedule the task. If the map percentage stays at 100% in the logs, look at the CloudWatch metrics, specifically RunningMapTasks, to check whether the map task is still processing. You can also find this information using the Hadoop web interface on the master node.
If you are seeing this issue, there are several things you can try:
Instruct the reduce phase to wait longer before starting. You can do this by altering the Hadoop configuration setting mapred.reduce.slowstart.completed.maps to a longer time. For more information, see Create Bootstrap Actions to Install Additional Software.
Match the reducer count to the total reducer capability of the cluster. You do this by adjusting the Hadoop configuration setting mapred.reduce.tasks for the job.
Use a combiner class code to minimize the amount of outputs that need to be fetched.
Check that there are no issues with the Amazon EC2 service that are affecting the network performance of the cluster. You can do this using the Service Health Dashboard.
Review the CPU and memory resources of the instances in your cluster to make sure that your data processing is not overwhelming the resources of your nodes. For more information, see Configure Cluster Hardware and Networking.
Check the version of the Amazon Machine Image (AMI) used in your Amazon EMR cluster. If the version is 2.3.0 through 2.4.4 inclusive, update to a later version. AMI versions in the specified range use a version of Jetty that may fail to deliver output from the map phase. The fetch error occurs when the reducers cannot obtain output from the map phase.
Jetty is an open-source HTTP server that is used for machine to machine communications within a Hadoop cluster
Does it have any measurable effect on resources whether I submit a bunch of hadoop jobs from different client servers or all from the same one? I would think not since all the work is done in the cluster. Is this correct?
The only thing which is resource intensive on the client submitting to the Hadoop cluster is the calculation of the input splits. When the input data is huge or when too many jobs are submitted from the same client then because of the input split calculations, the job submission might become a bit slow.
I am not able to recall the Hadoop release or the parameter, but a configurable parameter was included to move the calculation of the input splits from the client submitting a job to the Hadoop cluster.
It really shouldn't matter where you submit your jobs from. The client itself doesn't do much, it uses RPC protocol to contact the services, and then just sits idle until the job is finished.
Also, the most important is what kind of scheduler you use to allocate resource, which is probably going to make the most significant difference and decide which resources to allocate to which job. More on job scheduling here.
I don't think you can move the input split calculation into Job Tracker in 'Classic' version. In YARN, you can move it using
"yarn.app.mapreduce.am.compute-splits-in-cluster"
I am guessing, Hadoop people didn't want to overload Job tracker with input split creation. Similar to the design decision of not assigning too much work for Namenode in HDFS.
In YARN, every job gets its own Application Master, so no worries about overloading a SPOF/bottleneck master like job tracker.
In reference to the original question, the client job would have to reach out to the namenode to get the block locations (I have see parts of code on block storage class calling data node for some meta data...not sure whether these happen during input split creation or in task tracker node) . This can become an issue if you are handling a lot of jobs on the same client node.
If you are using YARN, there would be a slight performance increase if all these communications happen inside the cluster.
Need to check how Oozie handles this issue.
Hopefully, this helps!
Arun