Spark nullio File Format

Spark nullio file format is a no-op file format which is used for testing and benchmarking I/O and software cost of Spark input/output routines. What does it mean?

• For output: all data written to nullio file format is discarded.
• For input: it can selectively generate particular schema data on the fly. See Generating Input Data section for more details.

Updates

• September 21, 2017: The build is moved from Spark 2.1 to 2.2.
• September 21, 2017: Package rename from com.ibm.crail.spark.sql.datasources to org.apache.spark.sql
• August 10, 2017: The initial source code release for Spark 2.1.

How to compile

You can simply clone and compile it as

 #!/bin/bash 
 set -e 
 MAVEN_OPTS="-XX:+TieredCompilation -XX:TieredStopAtLevel=1"
 mvn -DskipTests -T 1C  package

If successful, this will give you spark-nullio-1.0.jar in the target folder. There are two ways to tell Spark about this jar

(1) copy this jar into your Spark's jar folder. (2) copy this jar to a location (lets say /tmp/spark-nullio-1.0.jar) and add the location to the spark config (conf/spark-defaults.conf) as

spark.driver.extraClassPath     /tmp/*
spark.executor.extraClassPath   /tmp/*

How to use Spark nullio file format

Most common usecase of nullio is to decouple storage devices from the rest of the Spark I/O code path, and isolate performance and debugging issues in the latter. For example, by measuring the runtime of a benchmark that writes to a stable storage (lets says HDFS) and one that just discards (using the nullio format), one can calculate how much time is spent in the actual data writing.

For discarding output

The general high-level syntax to write a Dataset in Spark (uses DataFrameWriter class[1] ) is

scala> daatset.write.format("format").options(...).mode(savemode).save("fileName")

For example, to write as a parquet file (and overwrite old data [2]) this would look like

scala> daatset.write.format("parquet").mode(SaveMode.Overwrite).save("/example.parquet")

Since, nullio is not a built in format for Spark, you have to specify the complete class path for format as

scala> daatset.write.format("org.apache.spark.sql.NullFileFormat").save("/example.nullio")

and that is it. This way, whatever data was in dataset will be pushed to the nullio writer, that will eventually discard all the data. When the writer is closed, it will print some statistics. You can check them in the driver or executor's log.

[1] https://spark.apache.org/docs/latest/api/scala/index.html#org.apache.spark.sql.DataFrameWriter

[2] https://spark.apache.org/docs/latest/sql-programming-guide.html#save-modes

For generating input data

Generating input data is a bit more involved. Here you can configure the nullio reader to generate specific type of schema data on the fly. There are some schema which are pre-coded, but you can add whatever you like as shown in the next section.

The general high-level syntax to read input source into a Dataset (sues DataFrameReader[3]) looks something like:

scala>spark.read.format("format").options(...).load("filename")

The input path can be configured with following parameters (you can pass them in options):

• inputrows : Number of input rows per task (default: 1000)
• payloadsize : size of any variable payload (default: 32 bytes)
• intrange : the range of the integers to control the data distribution. (default: INT_MAX)
• schema : the data schema. There are currently 3 options: (i) IntWithPayload <intKey:Int, payload:Array[Byte]>; (ii) ParquetExample<intKey:Int, randLong: Long, randDouble: Double, randFloat: Float, randString: String>; (iii) NullSchema (NYI).

Current limitation: Currently, I do not have a sensible way to control the number of partitions/splits. Hence, the current code expects that you will pass a sensible "filename". The rest of the spark mechaniary use that calculate number of tasks to launch. So for example, if reading your parquet file generated 83 tasks, then just changing the format type to NullFileFormat will still generate 83 tasks but the generated data will be what you set in the options. (see above)

How to add your own input schema and options

In order to add your schema, please have a look into ./schema/ examples. You just have to define number of fields, schema, and convert your data into InternalRows using UnsafeWriters.

Example usage

Reading data from a parquet file and discarding it

scala> val inputData = spark.read.format("parquet").load("/sampleParquet")
...
inputdata: org.apache.spark.sql.DataFrame = ...
scala> inputData.write.format("org.apache.spark.sql.NullFileFormat").save("/empty") 
...

Here in the executor log's (or in the dirver's log, depending upon your setting) you might see a line like this

closing Spark(Nullio)OutputWriter. init-to-write: 489947.3 usec  runTime: 425123.88 usec  #InternalRow: 1440202 #Rows: 0 time/Internalrow: 295 nsec time/Row: 0 nsec

This shows that this particular writer was asked to write 1440202 InternalRows which it discarded.

Reading data from nullio reader and saving it

We will generate 1001 rows (per executor) of IntWithPayload schema with 63 bytes of payload.

scala> val inputData = spark.read.format("org.apache.spark.sql.NullFileFormat")
   .option("inputrows", "1001")
   .option("payloadsize", "63")
   .option("schema","intwithpayload")
   .load("/sampleParquet")
   
inputData: org.apache.spark.sql.DataFrame = [intKey: int, payload: binary]
scala> inputData.write.parquet.save("/abc.parquet")
...
scala> spark.read.parquet("file:/data/abc").count 
...
res0: Long = 2002

2002 is good, as I have 2 executors.

Contributions

PRs are always welcome. Please fork, and make necessary modifications you propose, and let us know.

Contact

If you have questions or suggestions, feel free to post at:

https://groups.google.com/forum/#!forum/zrlio-users

or email: zrlio-users@googlegroups.com