Hi Jiang, Please refer to http://sortbenchmark.org/. When you take a look at the specification of each node Spark team uses, you can easily realize that # of nodes is not the only thing to take into consideration. You miss important things to consider for a fair comparison.
(1) # of disks in each node and disk type - Spark team has 8 SDD's in each node while we have 6 HDDs in each node. - If a SSD which Spark team uses can read data 500MB/sec, 8 SSD's can read 4000MB/sec. - If a HDD which I use can read data 100MB/sec, 6 HDD's can read 600MB/sec. - Here Spark team has 6.6x throughput. - Additionally, SSD's can read/write much better than HDDs with random disk access pattern. (2) the size of memory in each node - They have 244GB in each node while we have 24GB in each node. - The Spark team processes 494.7GB per node (100TB/207nodes) which is around 2x than 244GB (memory in each node). - As you know, I did sorting on 80GB per node which is around 3.5x than 24GB (memory in each node). (Actually I only use 12GB of memory in each node, so I process 7x data than used memory in each node.) - As you can see the following link, Spark gets slower when it processes more data than its memory: 35 page in http://www.slideshare.net/KostasTzoumas/apache-flink-api-runtime-and-project-roadmap . (3) # of CPU cores : they have 32 cores in each node while we have 16 cores in each node. The throughput you mentioned cannot be explained in that simple way. Regards, Dongwon Kim Postdoctoral Researcher @ Postech 2015-07-13 8:00 GMT+09:00 Hawin Jiang <hawin.ji...@gmail.com>: > Hi Kim and Stephan > > Kim's report is sorting 3360GB per 1427 seconds by Flink 0.9.0. 3360 > = 80*42 ((80GB/per node and 42 nodes) > Based on Kim's report. The TPS is 2.35GB/sec for Flink 0.9.0 > > Kim was using 42 nodes for testing purposes. I found that the best Spark > performance result was using 190 nodes from Databricks > The scalability factor is 42:190. if we are going to use 190 nodes for > this testing. > The Flink TPS should be 10.65GB /sec > > > Here is the summary table for your reference. > Please let me know if you have any questions about this table. > Thanks. > > [image: Inline image 2] > > > 72.93GB/sec = (1000TB*1024) / (234min*60) > > > The performance test report from Databricks. > > https://databricks.com/blog/2014/11/05/spark-officially-sets-a-new-record-in-large-scale-sorting.html > > > > > > > Best regards > Hawin > > On Fri, Jul 10, 2015 at 1:33 AM, Stephan Ewen <se...@apache.org> wrote: > >> Hi Dongwon Kim! >> >> Thank you for trying out these changes. >> >> The OptimizedText can be sorted more efficiently, because it generates a >> binary key prefix. That way, the sorting needs to serialize/deserialize >> less and saves on CPU. >> >> In parts of the program, the CPU is then less of a bottleneck and the >> disks and the network can unfold their bandwidth better. >> >> Greetings, >> Stephan >> >> >> >> On Fri, Jul 10, 2015 at 9:35 AM, Dongwon Kim <eastcirc...@postech.ac.kr> >> wrote: >> >>> Hi Stephan, >>> >>> I just pushed changes to my github: >>> https://github.com/eastcirclek/terasort. >>> I've modified the TeraSort program so that (A) it can reuse objects >>> and (B) it can exploit OptimizedText as you suggested. >>> >>> I've also conducted few experiments and the results are as follows: >>> ORIGINAL : 1714 >>> ORIGINAL+A : 1671 >>> ORIGINAL+B : 1467 >>> ORIGINAL+A+B : 1427 >>> Your advice works as shown above :-) >>> >>> Datasets are now defined as below: >>> - val inputFile = env.readHadoopFile(teraInputFormat, classOf[Text], >>> classOf[Text], inputPath) >>> - val optimizedText = inputFile.map(tp => (new OptimizedText(tp._1), >>> tp._2)) >>> - val sortedPartitioned = optimizedText.partitionCustom(partitioner, >>> 0).sortPartition(0, Order.ASCENDING) >>> - sortedPartitioned.map(tp => (tp._1.getText, tp._2)).output(hadoopOF) >>> You can see the two map transformations before and after the function >>> composition partitionCustom.sortPartition. >>> >>> Here is a question regarding the performance improvement. >>> Please see the attached Ganglia image files. >>> - ORIGINAL-[cpu, disks, network].png are for ORIGINAL. >>> - BEST-[cpu, disks, network].png are for ORIGINAL+A+B. >>> Compared to ORIGINAL, BEST shows better utilization of disks and >>> network and shows lower CPU utilization. >>> Is this because OptimizedText objects are serialized into Flink memory >>> layout? >>> What happens when keys are represented in just Text, not >>> OptimziedText? Are there another memory area to hold such objects? or >>> are they serialized anyway but in an inefficient way? >>> If latter, is the CPU utilization in ORIGINAL high because CPUs work >>> hard to serialize Text objects using Java serialization mechanism with >>> DataInput and DataOutput? >>> If true, I can explain the high throughput of network and disks in >>> ORIGINAL+A+B. >>> I, however, observed the similar performance when I do mapping not >>> only on 10-byte keys but also on 90-byte values, which cannot be >>> explained by the above conjecture. >>> Could you make things clear? If so, I would be very appreciated ;-) >>> >>> I'm also wondering whether the two map transformations, >>> (Text, Text) to (OptimizedText, Text) and (OptimizedText, Text) to >>> (Text, Text), >>> can prevent Flink from performing a lot better. >>> I don't have time to modify TeraInputFormat and TeraOutputFormat to >>> read (String, String) pairs from HDFS and write (String, String) pairs >>> to HDFS. >>> Do you see that one can get a better TeraSort result using an new >>> implementation of FileInputFormat<String,String>? >>> >>> Regards, >>> >>> Dongwon Kim >>> >>> 2015-07-03 3:29 GMT+09:00 Stephan Ewen <se...@apache.org>: >>> > Hello Dongwon Kim! >>> > >>> > Thanks you for sharing these numbers with us. >>> > >>> > I have gone through your implementation and there are two things you >>> could >>> > try: >>> > >>> > 1) >>> > >>> > I see that you sort Hadoop's Text data type with Flink. I think this >>> may be >>> > less efficient than if you sort String, or a Flink specific data type. >>> > >>> > For efficient byte operations on managed memory, Flink needs to >>> understand >>> > the binary representation of the data type. Flink understands that for >>> > "String" and many other types, but not for "Text". >>> > >>> > There are two things you can do: >>> > - First, try what happens if you map the Hadoop Text type to a Java >>> String >>> > (only for the tera key). >>> > - Second, you can try what happens if you wrap the Hadoop Text type >>> in a >>> > Flink type that supports optimized binary sorting. I have pasted code >>> for >>> > that at the bottom of this email. >>> > >>> > 2) >>> > >>> > You can see if it helps performance if you enable object re-use in >>> Flink. >>> > You can do this on the ExecutionEnvironment via >>> > "env.getConfig().enableObjectReuse()". Then Flink tries to use the same >>> > objects repeatedly, in case they are mutable. >>> > >>> > >>> > Can you try these options out and see how they affect Flink's runtime? >>> > >>> > >>> > Greetings, >>> > Stephan >>> > >>> > --------------------------------------------------------- >>> > Code for optimized sortable (Java): >>> > >>> > public final class OptimizedText implements >>> NormalizableKey<OptimizedText > >>> > { >>> > private final Text text; >>> > public OptimizedText () { >>> > this.text = new Text(); >>> > } >>> > public OptimizedText (Text from) { >>> > this.text = from; >>> > } >>> > >>> > public Text getText() { >>> > return text; >>> > } >>> > >>> > @Override >>> > public int getMaxNormalizedKeyLen() { >>> > return 10; >>> > } >>> > >>> > @Override >>> > public void copyNormalizedKey(MemorySegment memory, int offset, int >>> len) { >>> > memory.put(offset, text.getBytes(), 0, Math.min(text.getLength(), >>> > Math.min(10, len))); >>> > } >>> > >>> > @Override >>> > public void write(DataOutputView out) throws IOException { >>> > text.write(out); >>> > } >>> > >>> > @Override >>> > public void read(DataInputView in) throws IOException { >>> > text.readFields(in); >>> > } >>> > >>> > @Override >>> > public int compareTo(OptimizedText o) { >>> > return this.text.compareTo(o.text); >>> > } >>> > } >>> > >>> > --------------------------------------------------------- >>> > Converting Text to OptimizedText (Java code) >>> > >>> > map(new MapFunction<Tuple2<Text, Text>, Tuple2<OptimizedText, Text>>() >>> { >>> > @Override >>> > public Tuple2<OptimizedText, Text> map(Tuple2<Text, Text> value) { >>> > return new Tuple2<OptimizedText, Text>(new OptimizedText(value.f0), >>> > value.f1); >>> > } >>> > }) >>> > >>> > >>> > >>> > >>> > On Thu, Jul 2, 2015 at 6:47 PM, Dongwon Kim <eastcirc...@postech.ac.kr >>> > >>> > wrote: >>> >> >>> >> Hello, >>> >> >>> >> I'd like to share my code for TeraSort on Flink and Spark which uses >>> >> the same range partitioner as Hadoop TeraSort: >>> >> https://github.com/eastcirclek/terasort >>> >> >>> >> I also write a short report on it: >>> >> >>> >> >>> http://eastcirclek.blogspot.kr/2015/06/terasort-for-spark-and-flink-with-range.html >>> >> In the blog post, I make a simple performance comparison between >>> >> Flink, Spark, Tez, and MapReduce. >>> >> >>> >> I hope it will be helpful to you guys! >>> >> Thanks. >>> >> >>> >> Dongwon Kim >>> >> Postdoctoral Researcher @ Postech >>> > >>> > >>> >> >> >
