spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From kiran lonikar <loni...@gmail.com>
Subject Re: Optimisation advice for Avro->Parquet merge job
Date Mon, 08 Jun 2015 07:00:00 GMT
James,

As I can see, there are three distinct parts to your program:

   - for loop
   - synchronized block
   - final outputFrame.save statement

Can you do a separate timing measurement by putting a simple
System.currentTimeMillis() around these blocks to know how much they are
taking and then try to optimize where it takes longest? In the second
block, you may want to measure the time for the two statements. Improving
this boils down to playing with spark settings.

Now consider the first block: I think this is a classic case of merge sort
or a reduce tree. You already tried to improve this by submitting jobs in
parallel using executor pool/Callable etc.

To further improve the parallelization, I suggest you use a reduce tree
like approach. For example, lets say you want to compute sum of all
elements of an array in parallel. The way its solved for a GPU like
platform is you divide your input array initially in chunks of 2, compute
those n/2 sums parallely on separate threads and save the result in the
first of the two elements. In the next iteration, you compute n/4 sums
parallely of the earlier sums and so on till you are left with only two
elements whose sum gives you final sum.

You are performing many sequential unionAll operations for inputs.size()
avro files. Assuming the unionAll() on DataFrame is blocking (and not a
simple transformation like on RDDs) and actually performs the union
operation, you will certainly benefit by parallelizing this loop. You may
change the loop to something like below:

// pseudo code only
int n = inputs.size()
// initialize executor
executor = new FixedThreadPoolExecutor(n/2)
dfInput = new DataFrame[n/2]
for(int i =0;i < n/2;i++) {
    executor.submit(new Runnable() {
        public void run() {
            // union of i and i+n/2
            // showing [] only to bring out array access. Replace with
dfInput(i) in your code
            dfInput[i] = sqlContext.load(inputPaths.get(i),
"com.databricks.spark.avro").unionAll(sqlContext.load(inputsPath.get(i +
n/2), "com.databricks.spark.avro"))
        }
    });
}

executor.awaitTermination(0, TimeUnit.SECONDS)

int steps = log(n)/log(2.0)
for(s = 2; s < steps;s++) {
    int stride = n/(1 << s); // n/(2^s)
    for(int i = 0;i < stride;i++) {
        executor.submit(new Runnable() {
            public void run() {
                // union of i and i+n/2
                // showing [] only to bring out array access. Replace with
dfInput(i) and dfInput(i+stride) in your code
                dfInput[i] = dfInput[i].unionAll(dfInput[i + stride])
            }
        });
    }
    executor.awaitTermination(0, TimeUnit.SECONDS)
}

Let me know if it helped.

-Kiran


On Thu, Jun 4, 2015 at 8:57 PM, James Aley <james.aley@swiftkey.com> wrote:

> Thanks for the confirmation! We're quite new to Spark, so a little
> reassurance is a good thing to have sometimes :-)
>
> The thing that's concerning me at the moment is that my job doesn't seem
> to run any faster with more compute resources added to the cluster, and
> this is proving a little tricky to debug. There are a lot of variables, so
> here's what we've tried already and the apparent impact. If anyone has any
> further suggestions, we'd love to hear!
>
> * Increase the "minimum" number of output files (targetPartitions above),
> so that input groups smaller than our minimum chunk size can still be
> worked on by more than one executor. This does measurably speed things up,
> but obviously it's a trade-off, as the original goal for this job is to
> merge our data into fewer, larger files.
>
> * Submit many jobs in parallel, by running the above code in a Callable,
> on an executor pool. This seems to help, to some extent, but I'm not sure
> what else needs to be configured alongside it -- driver threads, scheduling
> policy, etc. We set scheduling to "FAIR" when doing this, as that seemed
> like the right approach, but we're not 100% confident. It seemed to help
> quite substantially anyway, so perhaps this just needs further tuning?
>
> * Increasing executors, RAM, etc. This doesn't make a difference by itself
> for this job, so I'm thinking we're already not fully utilising the
> resources we have in a smaller cluster.
>
> Again, any recommendations appreciated. Thanks for the help!
>
>
> James.
>
> On 4 June 2015 at 15:00, Eugen Cepoi <cepoi.eugen@gmail.com> wrote:
>
>> Hi
>>
>> 2015-06-04 15:29 GMT+02:00 James Aley <james.aley@swiftkey.com>:
>>
>>> Hi,
>>>
>>> We have a load of Avro data coming into our data systems in the form of
>>> relatively small files, which we're merging into larger Parquet files with
>>> Spark. I've been following the docs and the approach I'm taking seemed
>>> fairly obvious, and pleasingly simple, but I'm wondering if perhaps it's
>>> not the most optimal approach.
>>>
>>> I was wondering if anyone on this list might have some advice to make to
>>> make this job as efficient as possible. Here's some code:
>>>
>>> DataFrame dfInput = sqlContext.load(inputPaths.get(0),
>>> "com.databricks.spark.avro");
>>> long totalSize = getDirSize(inputPaths.get(0));
>>>
>>> for (int i = 1; i < inputs.size(); ++i) {
>>>     dfInput = dfInput.unionAll(sqlContext.load(inputPaths.get(i),
>>> "com.databricks.spark.avro"));
>>>     totalSize += getDirSize(inputPaths.get(i));
>>> }
>>>
>>> int targetPartitions = (int) Math.max(2L, totalSize / TARGET_SIZE_BYTES);
>>> DataFrame outputFrame;
>>>
>>> // Note: HADOOP-10456 impacts us, as we're stuck on 2.4.0 in EMR, hence
>>> // the synchronize block below. Suggestions welcome here too! :-)
>>> synchronized (this) {
>>>     RDD<Row> inputRdd = dfInput.rdd().coalesce(targetPartitions, false,
>>> null);
>>>     outputFrame = sqlContext.createDataFrame(inputRdd, dfInput.schema());
>>> }
>>>
>>> outputFrame.save(outputPath, "parquet", SaveMode.Overwrite);
>>>
>>> Here are some things bothering me:
>>>
>>>    - Conversion to an RDD and back again so that we can use coalesce()
>>>    to reduce the number of partitions. This is because we read that
>>>    repartition() is not as efficient as coalesce(), and local micro benchmarks
>>>    seemed to somewhat confirm that this was faster. Is this really a good idea
>>>    though? Should we be doing something else?
>>>
>>> Repartition uses coalesce but with a forced shuffle step. Its just a
>> shortcut for coalesce(xxx, true)
>> Doing a coalesce sounds correct, I'd do the same :) Note that if you add
>> the shuffle step, then your partitions should be better balanced.
>>
>>>
>>>    - Usage of unionAll() - this is the only way I could find to join
>>>    the separate data sets into a single data frame to save as Parquet. Is
>>>    there a better way?
>>>
>>> When using directly the inputformats you can do this
>> FileInputFormat.addInputPath, it should perform at least as good as union.
>>
>>>
>>>    - Do I need to be using the DataFrame API at all? I'm not querying
>>>    any data here, so the nice API for SQL-like transformations of the data
>>>    isn't being used. The DataFrame API just seemed like the path of least
>>>    resistance for working with Avro and Parquet. Would there be any advantage
>>>    to using hadoopRDD() with the appropriate Input/Output formats?
>>>
>>>
>>>
>> Using directly the input/outputformats sounds viable. But the snippet you
>> show seems clean enough and I am not sure there would be much value in
>> making something (maybe) slightly faster but harder to understand.
>>
>>
>> Eugen
>>
>> Any advice or tips greatly appreciated!
>>>
>>>
>>> James.
>>>
>>>
>>>
>>
>

Mime
View raw message