Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
SPARK-2047: Introduce an in-mem Sorter, and use it to reduce mem usage
### Why and what? Currently, the AppendOnlyMap performs an "in-place" sort by converting its array of [key, value, key, value] pairs into a an array of [(key, value), (key, value)] pairs. However, this causes us to allocate many Tuple2 objects, which come at a nontrivial overhead. This patch adds a Sorter API, intended for in memory sorts, which simply ports the Android Timsort implementation (available under Apache v2) and abstracts the interface in a way which introduces no more than 1 virtual function invocation of overhead at each abstraction point. Please compare our port of the Android Timsort sort with the original implementation: http://www.diffchecker.com/wiwrykcl ### Memory implications An AppendOnlyMap contains N kv pairs, which results in roughly 2N elements within its underlying array. Each of these elements is 4 bytes wide in a [compressed OOPS](https://wikis.oracle.com/display/HotSpotInternals/CompressedOops) system, which is the default. Today's approach immediately allocates N Tuple2 objects, which take up 24N bytes in total (exposed via YourKit), and undergoes a Java sort. The Java 6 version immediately copies the entire array (4N bytes here), while the Java 7 version has a worst-case allocation of half the array (2N bytes). This results in a worst-case sorting overhead of 24N + 2N = 26N bytes (for Java 7). The Sorter does not require allocating any tuples, but since it uses Timsort, it may copy up to half the entire array in the worst case. This results in a worst-case sorting overhead of 4N bytes. Thus, we have reduced the worst-case overhead of the sort by roughly 22 bytes times the number of elements. ### Performance implications As the destructiveSortedIterator is used for spilling in an ExternalAppendOnlyMap, the purpose of this patch is to provide stability by reducing memory usage rather than improve performance. However, because it implements Timsort, it also brings a substantial performance boost over our prior implementation. Here are the results of a microbenchmark that sorted 25 million, randomly distributed (Float, Int) pairs. The Java Arrays.sort() tests were run **only on the keys**, and thus moved less data. Our current implementation is called "Tuple-sort using Arrays.sort()" while the new implementation is "KV-array using Sorter". <table> <tr><th>Test</th><th>First run (JDK6)</th><th>Average of 10 (JDK6)</th><th>First run (JDK7)</th><th>Average of 10 (JDK7)</th></tr> <tr><td>primitive Arrays.sort()</td><td>3216 ms</td><td>1190 ms</td><td>2724 ms</td><td>131 ms (!!)</td></tr> <tr><td>Arrays.sort()</td><td>18564 ms</td><td>2006 ms</td><td>13201 ms</td><td>878 ms</td></tr> <tr><td>Tuple-sort using Arrays.sort()</td><td>31813 ms</td><td>3550 ms</td><td>20990 ms</td><td>1919 ms</td></tr> <tr><td><b>KV-array using Sorter</b></td><td></td><td></td><td><b>15020 ms</b></td><td><b>834 ms</b></td></tr> </table> The results show that this Sorter performs exactly as expected (after the first run) -- it is as fast as the Java 7 Arrays.sort() (which shares the same algorithm), but is significantly faster than the Tuple-sort on Java 6 or 7. In short, this patch should significantly improve performance for users running either Java 6 or 7. Author: Aaron Davidson <aaron@databricks.com> Closes apache#1502 from aarondav/sort and squashes the following commits: 652d936 [Aaron Davidson] Update license, move Sorter to java src a7b5b1c [Aaron Davidson] fix licenses 5c0efaf [Aaron Davidson] Update tmpLength ec395c8 [Aaron Davidson] Ignore benchmark (again) and fix docs 034bf10 [Aaron Davidson] Change to Apache v2 Timsort b97296c [Aaron Davidson] Don't try to run benchmark on Jenkins + private[spark] 6307338 [Aaron Davidson] SPARK-2047: Introduce an in-mem Sorter, and use it to reduce mem usage
- Loading branch information
