Skip to main content



Spark - Distributed Computing and Multi-Threading


Spark on Hadoop works on concept of data locality, such that code is tried to send to data rather then shuffling data to code.


We specify following resources when creating a Spark Job - 

  • executor cores
  • executor memory
  • driver cores
  • driver memory
  • number of executors


Above property in general determines number of executors and each executor is kind of JVM process with specified cores and memory to use. The executors connects back to your driver program. Now the driver can send them commands. 

Data is distributed on HDFS in blocks, and an input split kind of determine size of a data partition. YARN tries to collocate executors as close to data as possible to minimize network transfer of data.

A task is a command sent from the driver to an executor by serializing your Function object. The executor deserializes the command, and executes it on a partition.

The number of the Spark tasks in a single stage equals to the number of RDD partitions.

  • Now, "executor cores" or "spark.executor.cores" determines number of tasks that can be executed concurrently on an executor. 
  • What is a task ? As explained above, task is a function that needs to be executed on a partition of data. What function? function as defined by map(), flatMap(), forEach(), etc.
So, based of "executor cores" and "number of executors" - Spark does distributed computing by executing that many tasks in parallel.

Now, one may further use ThreadPools or Streams within a task to do quick computation/ processing. In this scenario, we can configure "spark.task.cpus" to determine number of cores to allocate to each task. Make sure to use Java Concurrent API and proper Synchronization to avoid misleading results. That's how one can achieve Multi-Threading within a task.


A command-line utility “lscpu” in Linux is used to get CPU information of the system. The “lscpu” command fetches the CPU architecture information from the “sysfs” and /proc/cpuinfo files and displays it in a terminal. What matters to you is - 
  • Thread(s) per core (t)
  • Core(s) per socket (c)
  • Socket (s)
Number of Logical Cores/ Total Number of threads  =  t*c*s
Number of Cores = s*c

So, say if you have configuration like below - 
Thread(s) per core:    2
Core(s) per socket:    12

Socket(s):             4 

And, you want Multi-Threading by 4 threads. Then set "spark.task.cpus=2", as each CPU can have 2 threads, meaning 2*2 = 4 threads.

Comments

Post a Comment

Popular posts

Unix Server ( Edge Node ) hangs when there are many jobs running on hadoop cluster started from Unix Edge Node.

  When a unix server or an edge node is running lots of jobs (like Spark, Hadoop, or custom batch processes), crashes happen. For example. For example a process might hit a segementation fault, memory issue or ay other runtime issue. By default, if ulimit -c is not 0, the OS will create core dump. Core dump are written to disk and can be very large, sometimes hundreds of MBs or even GBs per process. What we realized was that when multiple processes crash at the same time, the system suddenly tries to write core files to disk. This was leading to DisK I/O spikes. Thus, node was becoming unresponsive. This was also leading CPU spike because OS was handling crash logging. Setting "ulimit -c 0" disables core dumps. This way we lose ability to debug crashes via core dump But, kept production edge nodes stable. On most Linux systems, by default, "core dumps" are written in current working directory of the process that crashes. Linux allows you to change core dump file nam...
Post a Comment
Read more




Spring MongoDB Rest API not returning response in 90 seconds which is leading to client timeout

  We have Spring Boot  Rest API deployed in Kubernetes cluster which integrates with MongoDB to fetch the data.  MongoDB is fed with data by a real time Spark & NiFi job.  Our clients complained that for a request what they send they don't have response within 90 seconds. Consider it like an OMS ( Order ManagEment System).  On further analysis, we found that Spark & NiFi processing is happenning within 10 seconds after consuming response data from Kafka. Thus, initally out thought was that it due to delay from upstream to produce data in to Kafka.  Thankfully, our data had create / request  timestamp, and when response was received, and when response was inserted into MongoDB. Subtracting response insert time from request time seemed to be well within 90 seconds. But, still client did timeout on not seeing a response within 90 seconds. This led to confusion on our side.  But, then we realized it was due to Read Preference . We updated this...
Post a Comment
Read more




MongoDB Regex Query taking more time in Production but same query perform well in UAT

   We came across a situation where-in, MongoDB Query was taking more time in Production like 10 seconds and 4.2 seconds but same query performed well in UAT taking under 400 ms. The very first thought that was evident to us that it is because of amount of data which differed in UAT and Production. Then we ran following to see the execution plan -   db.collection.aggregate(<queries>).explain() This gave us Winning and Rejected Plans. Under which, we analyzed that although it was using 'IXSCAN.' But, it was incorrect index- as we had one compound index built on time field and other fields, and there was other index just on time field for TTL purposes. Winning plan picked TTL index rather than compound index. Thus, we dropped TTL index and built TTL index on a different time field.  That got our query performance time from 10 seconds to 726 ms. Also, for other query the performance came down from 8 seconds to 4.3 seconds. Then, we ran following -  ...
Post a Comment
Read more




Machine Learning Part 1

Machine learning uses algorithms to find patterns in data, and then uses a model that recognizes those patterns to make predictions on new data. Machine learning may be broken down into - Supervised learning algorithms use labeled data - Classification, Regression. Unsupervised learning algorithms find patterns in unlabeled data - Clustering, Collaborative Filtering, Frequent Pattern Mining Semi-supervised learning uses a mixture of labeled and unlabeled data. Reinforcement learning trains algorithms to maximize rewards based on feedback. Classification - Mailing Servers like Gmail uses ML to classify if an email is Spam or not based on the data of an email: the sender, recipients, subject, and message body. Classification takes a set of data with known labels and learns how to label new records based on that information. For example- An items is important or not. A transaction is fraud or not based upon known labeled examples of transactions which were classified...
Post a Comment
Read more




Clone multiple projects or repositories from GitLab under a folder or subfolder.

  Suppose we have a group named as groupA in gitlab, under which we have sub folder structure, and there are multiple projects under a folder.  We wish to clone all the projects under folder. This can be done using gitlab api. Like below -  1) You should have Token for authentication. TOKEN="J612a-xUoMxxcerssRe31_" 2) API URL to subfolder that should give you group id.  API_URL1="https://<gitlab.server.com>/api/v4/groups?search=groupA /dir1/dir2/dir3/dir4" 3) Use below to get group id - GROUP_ID=$(curl --silent --header "Private-Token: $TOKEN" "$API_URL1" | jq -r '.[].id') 4) Now you can fetch and clone all projects by using below - API_URL2="https://<gitlab.server.com>/api/v4/groups/$GROUP_ID/projects" # Fetch repositories list REPOS=$(curl --silent --header "Private-Token: $TOKEN" "$API_URL2" | jq -r '.[].http_url_to_repo') # Clone each repository for REPO in $REPOS; do     git clone $RE...
Post a Comment
Read more