Details
-
New Feature
-
Status: Resolved
-
Major
-
Resolution: Won't Fix
-
1.1.0
-
None
Description
I think a memory-based shuffle can reduce some overhead of disk I/O. I just want to know is there any plan to do something about it. Or any suggestion about it. Base on the work (SPARK-2044), it is feasible to have several implementations of shuffle.
----------------------------------------------------------------------------------------------------------------------------------------------------------------
Currently, there are two implementions of shuffle manager, i.e. SORT and HASH. Both of them will use disk in some stages. For examples, in the map side, all the intermediate data will be written into temporary files. In the reduce side, Spark will use external sort sometimes. In any case, disk I/O will bring some performance loss. Maybe，we can provide a pure-memory shuffle manager. In this shuffle manager, intermediate data will only go through memory. In some of scenes, it can improve performance. Experimentally, I implemented a in-memory shuffle manager upon SPARK-2044.
1. Following is my testing result (some heary shuffle operations):
| data size (Byte) | partitions | resources |
| 5131859218 | 2000 | 50 executors/ 4 cores/ 4GB |
| settings | operation1 | operation2 |
| shuffle spill & lz4 | repartition+flatMap+groupByKey | repartition + groupByKey |
| memory | 38s | 16s |
| sort | 45s | 28s |
| hash | 46s | 28s |
| no shuffle spill & lz4 | ||
| memory | 16s | 16s |
| shuffle spill & lzf | ||
| memory | 28s | 27s |
| sort | 29s | 29s |
| hash | 41s | 30s |
| no shuffle spill & lzf | ||
| memory | 15s | 16s |
In my implementation, I simply reused the "BlockManager" in the map-side and set the "spark.shuffle.spill" false in the reduce-side. All the intermediate data is cached in memory store. Just as Reynold Xin has pointed out, our disk-based shuffle manager has achieved a good performance. With parameter tuning, the disk-based shuffle manager will obtain similar performance as memory-based shuffle manager. However, I will continue my work and improve it. And as an alternative tuning option, "InMemory shuffle" is a good choice. Future work includes, but is not limited to:
- memory usage management in "InMemory Shuffle" mode
- data management when intermediate data can not fit in memory
Test code：
val conf = new SparkConf().setAppName("InMemoryShuffleTest")
val sc = new SparkContext(conf)
val dataPath = args(0)
val partitions = args(1).toInt
val rdd1 = sc.textFile(dataPath).cache()
rdd1.count()
val startTime = System.currentTimeMillis()
val rdd2 = rdd1.repartition(partitions)
.flatMap(_.split(",")).map(s => (s, s))
.groupBy(e => e._1)
rdd2.count()
val endTime = System.currentTimeMillis()
println("time: " + (endTime - startTime) / 1000 )
2. Following is a Spark Sort Benchmark (in spark 1.1.1). There is no tuning for disk shuffle.
2.1. Test the influence of memory size per core
precondition: 100GB(SORT benchmark), 100 executor /15cores 1491partitions (input file blocks) .
| memory size per executor | inmemory shuffle(no shuffle spill) | sort shuffle | hash shuffle | improvement(vs.sort) | improvement(vs.hash) |
| 9GB | 79.652849s | 60.102337s | failed | -32.7% | |
| 12GB | 54.821924s | 51.654897s | 109.167068s | -3.17% | +47.8% |
| 15GB | 33.537199s | 40.140621s | 48.088158s | +16.47% | +30.26% |
| 18GB | 30.930927s | 43.392401s | 49.830276s | +28.7% | +37.93% |
2.2. Test the influence of partition number
18GB/15cores per executor
| partitions | inmemory shuffle(no shuffle spill) | sort shuffle | hash shuffle | improvement(vs.sort) | improvement(vs.hash) |
| 1000 | 92.675436s | 85.193158s | 71.106323s | -8.78% | -30.34% |
| 1491 | 30.930927s | 43.392401s | 49.830276s | +28.7% | +37.93% |
| 2000 | 18.385s | 26.653720s | 30.103s | +31.02% | +38.92% |
Attachments
Issue Links
- is related to
-
-
- Resolved
-
- links to
