Apache Spark: The number of cores vs. the number o

问题:

I'm trying to understand the relationship of the number of cores and the number of executors when running a Spark job on YARN.

The test environment is as follows:

• Number of data nodes: 3
• Data node machine spec:
  • CPU: Core i7-4790 (# of cores: 4, # of threads: 8)
  • RAM: 32GB (8GB x 4)
  • HDD: 8TB (2TB x 4)

• Network: 1Gb

• Spark version: 1.0.0

• Hadoop version: 2.4.0 (Hortonworks HDP 2.1)

• Spark job flow: sc.textFile -> filter -> map -> filter -> mapToPair -> reduceByKey -> map -> saveAsTextFile

• Input data

  • Type: single text file
  • Size: 165GB
  • Number of lines: 454,568,833

• Output

  • Number of lines after second filter: 310,640,717
  • Number of lines of the result file: 99,848,268
  • Size of the result file: 41GB

The job was run with following configurations:

1. --master yarn-client --executor-memory 19G --executor-cores 7 --num-executors 3 (executors per data node, use as much as cores)

2. --master yarn-client --executor-memory 19G --executor-cores 4 --num-executors 3 (# of cores reduced)

3. --master yarn-client --executor-memory 4G --executor-cores 2 --num-executors 12 (less core, more executor)

Elapsed times:

1. 50 min 15 sec

2. 55 min 48 sec

3. 31 min 23 sec

To my surprise, (3) was much faster.
I thought that (1) would be faster, since there would be less inter-executor communication when shuffling.
Although # of cores of (1) is fewer than (3), #of cores is not the key factor since 2) did perform well.

(Followings were added after pwilmot's answer.)

For the information, the performance monitor screen capture is as follows:

• Ganglia data node summary for (1) - job started at 04:37.

• Ganglia data node summary for (3) - job started at 19:47. Please ignore the graph before that time.

The graph roughly divides into 2 sections:

• First: from start to reduceByKey: CPU intensive, no network activity
• Second: after reduceByKey: CPU lowers, network I/O is done.

As the graph shows, (1) can use as much CPU power as it was given. So, it might not be the problem of the number of the threads.

How to explain this result?

回答1:

To hopefully make all of this a little more concrete, here’s a worked example of configuring a Spark app to use as much of the cluster as possible: Imagine a cluster with six nodes running NodeManagers, each equipped with 16 cores and 64GB of memory. The NodeManager capacities, yarn.nodemanager.resource.memory-mb and yarn.nodemanager.resource.cpu-vcores, should probably be set to 63 * 1024 = 64512 (megabytes) and 15 respectively. We avoid allocating 100% of the resources to YARN containers because the node needs some resources to run the OS and Hadoop daemons. In this case, we leave a gigabyte and a core for these system processes. Cloudera Manager helps by accounting for these and configuring these YARN properties automatically.

The likely first impulse would be to use --num-executors 6 --executor-cores 15 --executor-memory 63G. However, this is the wrong approach because:

63GB + the executor memory overhead won’t fit within the 63GB capacity of the NodeManagers. The application master will take up a core on one of the nodes, meaning that there won’t be room for a 15-core executor on that node. 15 cores per executor can lead to bad HDFS I/O throughput.

A better option would be to use --num-executors 17 --executor-cores 5 --executor-memory 19G. Why?

This config results in three executors on all nodes except for the one with the AM, which will have two executors. --executor-memory was derived as (63/3 executors per node) = 21. 21 * 0.07 = 1.47. 21 – 1.47 ~ 19.

The explanation was given in a article in cloudera's blog http://blog.cloudera.com/blog/2015/03/how-to-tune-your-apache-spark-jobs-part-2/

回答2:

As you run your spark app on top of HDFS, according to Sandy Ryza

I’ve noticed that the HDFS client has trouble with tons of concurrent threads. A rough guess is that at most five tasks per executor can achieve full write throughput, so it’s good to keep the number of cores per executor below that number.

So I believe that your first configuration is slower than third one is because of bad HDFS I/O throughput

回答3:

I haven't played with these settings myself so this is just speculation but if we think about this issue as normal cores and threads in a distributed system then in your cluster you can use up to 12 cores (4 * 3 machines) and 24 threads (8 * 3 machines). In your first two examples you are giving your job a fair number of cores (potential computation space) but the number of threads (jobs) to run on those cores is so limited that you aren't able to use much of the processing power allocated and thus the job is slower even though there is more computation resources allocated.

you mention that your concern was in the shuffle step - while it is nice to limit the overhead in the shuffle step it is generally much more important to utilize the parallelization of the cluster. Think about the extreme case - a single threaded program with zero shuffle.

回答4:

From the excellent resources available at RStudio's Sparklyr package page:

SPARK DEFINITIONS:

It may be useful to provide some simple definitions for the Spark nomenclature:

Node: A server

Worker Node: A server that is part of the cluster and are available to run Spark jobs

Master Node: The server that coordinates the Worker nodes.

Executor: A sort of virtual machine inside a node. One Node can have multiple Executors.

Driver Node: The Node that initiates the Spark session. Typically, this will be the server where sparklyr is located.

Driver (Executor): The Driver Node will also show up in the Executor list.

回答5:

Short answer: I think tgbaggio is right. You hit HDFS throughput limits on your executors.

I think the answer here may be a little simpler than some of the recommendations here.

The clue for me is in the cluster network graph. For run 1 the utilization is steady at ~50 M bytes/s. For run 3 the steady utilization is doubled, around 100 M bytes/s.

From the cloudera blog post shared by DzOrd, you can see this important quote:

I’ve noticed that the HDFS client has trouble with tons of concurrent threads. A rough guess is that at most five tasks per executor can achieve full write throughput, so it’s good to keep the number of cores per executor below that number.

So, let's do a few calculations see what performance we expect if that is true.

---

Run 1: 19 GB, 7 cores, 3 executors

• 3 executors x 7 threads = 21 threads
• with 7 cores per executor, we expect limited IO to HDFS (maxes out at ~5 cores)
• effective throughput ~= 3 executors x 5 threads = 15 threads

Run 3: 4 GB, 2 cores, 12 executors

• 2 executors x 12 threads = 24 threads
• 2 cores per executor, so hdfs throughput is ok
• effective throughput ~= 12 executors x 2 threads = 24 threads

---

If the job is 100% limited by concurrency (the number of threads). We would expect runtime to be perfectly inversely correlated with the number of threads.

ratio_num_threads = nthread_job1 / nthread_job3 = 15/24 = 0.625
inv_ratio_runtime = 1/(duration_job1 / duration_job3) = 1/(50/31) = 31/50 = 0.62

So ratio_num_threads ~= inv_ratio_runtime, and it looks like we are network limited.

This same effect explains the difference between Run 1 and Run 2.

---

Run 2: 19 GB, 4 cores, 3 executors

• 3 executors x 4 threads = 12 threads
• with 4 cores per executor, ok IO to HDFS
• effective throughput ~= 3 executors x 4 threads = 12 threads

---

Comparing the number of effective threads and the runtime:

ratio_num_threads = nthread_job2 / nthread_job1 = 12/15 = 0.8
inv_ratio_runtime = 1/(duration_job2 / duration_job1) = 1/(55/50) = 50/55 = 0.91

It's not as perfect as the last comparison, but we still see a similar drop in performance when we lose threads.

Now for the last bit: why is it the case that we get better performance with more threads, esp. more threads than the number of CPUs?

A good explanation of the difference between parallelism (what we get by dividing up data onto multiple CPUs) and concurrency (what we get when we use multiple threads to do work on a single CPU) is provided in this great post by Rob Pike: Concurrency is not parallelism.

The short explanation is that if a Spark job is interacting with a file system or network the CPU spends a lot of time waiting on communication with those interfaces and not spending a lot of time actually "doing work". By giving those CPUs more than 1 task to work on at a time, they are spending less time waiting and more time working, and you see better performance.

回答6:

I think one of the major reasons is locality. Your input file size is 165G, the file's related blocks certainly distributed over multiple DataNodes, more executors can avoid network copy.

Try to set executor num equal blocks count, i think can be faster.

回答7:

Spark Dynamic allocation gives flexibility and allocates resources dynamically. In this number of min and max executors can be given. Also the number of executors that has to be launched at the starting of the application can also be given.

Read below on the same:

http://spark.apache.org/docs/latest/configuration.html#dynamic-allocation

回答8:

There is a small issue in the First two configurations i think. The concepts of threads and cores like follows. The concept of threading is if the cores are ideal then use that core to process the data. So the memory is not fully utilized in first two cases. If you want to bench mark this example choose the machines which has more than 10 cores on each machine. Then do the bench mark.

But dont give more than 5 cores per executor there will be bottle neck on i/o performance.

So the best machines to do this bench marking might be data nodes which have 10 cores.

Data node machine spec: CPU: Core i7-4790 (# of cores: 10, # of threads: 20) RAM: 32GB (8GB x 4) HDD: 8TB (2TB x 4)