Mailing-List: contact issues-help@drill.apache.org; run by ezmlm
Precedence: bulk
Reply-To: dev@drill.apache.org
Date: Tue, 22 Mar 2016 01:32:25 +0000 (UTC)
From: "ASF GitHub Bot (JIRA)" <jira@apache.org>
To: issues@drill.apache.org
Message-ID: <JIRA.12924985.1451528754000.12529.1458610345511@Atlassian.JIRA>
In-Reply-To: <JIRA.12924985.1451528754000@Atlassian.JIRA>
References: <JIRA.12924985.1451528754000@Atlassian.JIRA>
 <JIRA.12924985.1451528754337@arcas>
Subject: [jira] [Commented] (DRILL-4237) Skew in hash distribution
MIME-Version: 1.0
Content-Type: text/plain; charset=utf-8
Content-Transfer-Encoding: 7bit


    [ https://issues.apache.org/jira/browse/DRILL-4237?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15205589#comment-15205589 ] 

ASF GitHub Bot commented on DRILL-4237:
---------------------------------------

Github user jacques-n commented on a diff in the pull request:

    https://github.com/apache/drill/pull/430#discussion_r56925560
  
    --- Diff: exec/java-exec/src/main/java/org/apache/drill/exec/expr/fn/impl/HashHelper.java ---
    @@ -17,47 +17,77 @@
      */
     package org.apache.drill.exec.expr.fn.impl;
     
    +import io.netty.buffer.DrillBuf;
    +import org.apache.drill.common.config.DrillConfig;
    +import org.apache.drill.common.exceptions.DrillConfigurationException;
    +
     import java.nio.ByteBuffer;
     import java.nio.ByteOrder;
     
    -public class HashHelper {
    +public abstract class HashHelper {
       static final org.slf4j.Logger logger = org.slf4j.LoggerFactory.getLogger(HashHelper.class);
    +  public static final String defaultHashClassName = new String("org.apache.drill.exec.expr.fn.impl.MurmurHash3");
    +  static final String HASH_CLASS_PROP = "drill.exec.hash.class";
     
    +  static String actualHashClassName = defaultHashClassName;
    +  static DrillHash hashCall = new MurmurHash3();
    +  static {
     
    -  /** taken from mahout **/
    -  public static int hash(ByteBuffer buf, int seed) {
    -    // save byte order for later restoration
    -
    -    int m = 0x5bd1e995;
    -    int r = 24;
    +    try {
    +      DrillConfig config = DrillConfig.create();
    +      String configuredClassName = config.getString(HASH_CLASS_PROP);
    +      if(configuredClassName != null && configuredClassName != "") {
    +        actualHashClassName = configuredClassName;
    +        hashCall = config.getInstanceOf(HASH_CLASS_PROP, DrillHash.class);
    +      }
    +      logger.debug("HashHelper initializes with " + actualHashClassName);
    +    }
    +    catch(Exception ex){
    +      logger.error("Could not initialize Hash %s", ex.getMessage());
    +    }
    +  }
     
    -    int h = seed ^ buf.remaining();
    +  public static String getHashClassName(){
    +    return actualHashClassName;
    +  }
     
    -    while (buf.remaining() >= 4) {
    -      int k = buf.getInt();
    +  public static int hash32(int val, long seed) {
    +    double converted = val;
    +    return hash32(converted, seed);
    +  }
    +  public static int hash32(long val, long seed) {
    +    double converted = val;
    +    return hash32(converted, seed);
    +  }
    +  public static int hash32(float val, long seed){
    +    double converted = val;
    +    return hash32(converted, seed);
    +  }
     
    -      k *= m;
    -      k ^= k >>> r;
    -      k *= m;
    +  public static int hash32(double val, long seed){
    +    return hashCall.hash32(val, seed);
    +  }
     
    -      h *= m;
    -      h ^= k;
    -    }
    +  public static  int hash32(int start, int end, DrillBuf buffer, int seed){
    +    return hashCall.hash32(start, end, buffer, seed);
    --- End diff --
    
    Yes, I'm worried about the extra performance hit. I believe we already spend a reasonable amount of processing time applying hash functions and have considered it an opportunity for improvement. Give your current construction, we would need to dereference the field everytime we call the hash function. In the past my analysis of assembly out of the JVM is that this isn't typically removed. Directly binding to a static function doesn't require this overhead. Take a look at the jvm bytecode (or assembly) to see the difference. In general, our goal inside individual functions is to avoid indirection as much as possible, especially with a hot path such as the hash function. 


> Skew in hash distribution
> -------------------------
>
>                 Key: DRILL-4237
>                 URL: https://issues.apache.org/jira/browse/DRILL-4237
>             Project: Apache Drill
>          Issue Type: Bug
>          Components: Functions - Drill
>    Affects Versions: 1.4.0
>            Reporter: Aman Sinha
>            Assignee: Chunhui Shi
>
> Apparently, the fix in DRILL-4119 did not fully resolve the data skew issue.  It worked fine on the smaller sample of the data set but on another sample of the same data set, it still produces skewed values - see below the hash values which are all odd numbers. 
> {noformat}
> 0: jdbc:drill:zk=local> select columns[0], hash32(columns[0]) from `test.csv` limit 10;
> +-----------------------------------+--------------+
> |              EXPR$0               |    EXPR$1    |
> +-----------------------------------+--------------+
> | f71aaddec3316ae18d43cb1467e88a41  | 1506011089   |
> | 3f3a13bb45618542b5ac9d9536704d3a  | 1105719049   |
> | 6935afd0c693c67bba482cedb7a2919b  | -18137557    |
> | ca2a938d6d7e57bda40501578f98c2a8  | -1372666789  |
> | fab7f08402c8836563b0a5c94dbf0aec  | -1930778239  |
> | 9eb4620dcb68a84d17209da279236431  | -970026001   |
> | 16eed4a4e801b98550b4ff504242961e  | 356133757    |
> | a46f7935fea578ce61d8dd45bfbc2b3d  | -94010449    |
> | 7fdf5344536080c15deb2b5a2975a2b7  | -141361507   |
> | b82560a06e2e51b461c9fe134a8211bd  | -375376717   |
> +-----------------------------------+--------------+
> {noformat}
> This indicates an underlying issue with the XXHash64 java implementation, which is Drill's implementation of the C version.  One of the key difference as pointed out by [~jnadeau] was the use of unsigned int64 in the C version compared to the Java version which uses (signed) long.  I created an XXHash version using com.google.common.primitives.UnsignedLong.  However, UnsignedLong does not have bit-wise operations that are needed for XXHash such as rotateLeft(),  XOR etc.  One could write wrappers for these but at this point, the question is: should we think of an alternative hash function ? 
> The alternative approach could be the murmur hash for numeric data types that we were using earlier and the Mahout version of hash function for string types (https://github.com/apache/drill/blob/master/exec/java-exec/src/main/java/org/apache/drill/exec/expr/fn/impl/HashHelper.java#L28).  As a test, I reverted to this function and was getting good hash distribution for the test data. 
> I could not find any performance comparisons of our perf tests (TPC-H or DS) with the original and newer (XXHash) hash functions.  If performance is comparable, should we revert to the original function ?  
> As an aside, I would like to remove the hash64 versions of the functions since these are not used anywhere. 


--
This message was sent by Atlassian JIRA
(v6.3.4#6332)