cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "sankalp kohli (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (CASSANDRA-4310) Make Level Compaction go faster(multiple independant compactions in parallel) for insert heavy workload
Date Sat, 09 Jun 2012 00:38:23 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-4310?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13292129#comment-13292129
] 

sankalp kohli commented on CASSANDRA-4310:
------------------------------------------

One of the ways of speeding up level compaction mostly when using SSDs is to do them in parallel.
Here is the algorithm.

public List<Set<sstableReader>> getIndependantCompactions() {
Create a set of stables called compacting. 
Create a List of Set of stables like List<Set<sstableReader>> compactionList.
LIMIT_OF_COMPACTIONS = the number of compactions to do in parallel; 

//Ignore L0 for now. 
for(int level = MAX(8) to 1) {
  while(level is eligible and compactionList.size < LIMIT_OF_COMPACTIONS and we have not
checked for all stables in this level) {
     select the next stable from level L and also its overlapping stables in L+1 level like
we do now.
     if stables found in previous step do not overlap with any compacting stable(use compacting
set to store the stables which are compacting) in this set, add then to compactionList. 
     persist next for next call to this method like we do now so that when this level become
eligible again, it continues from here. 
    
  }
   
  if(compactionList.size >= LIMIT_OF_COMPACTIONS)
     return compactionList;

  //DO it for L0
  //Level 0 is a little different as stables don't don't have mutual exclusion of keys. 
  //So we have to check whether two overlapping stables are not schedules as two different
compactions in parallel.  
  Create a set of stables called L0selected - This will store the stables used for compactions.
We will also use compacting. 
  while(L0 is eligible and compactionList.size < LIMIT_OF_COMPACTIONS and we have not checked
for all stables in this level) {
     Find the oldest stable(like normal) which is not being compacted(by checking in compacting
set) and does not overlap with any stable in L0selected.
     Find up to 31 overlapping stables with the stable found in previous step in L0.
     if none of these 32(31 + 1) stables overlap with any stable in L0Selected set or are
in it then do next step
            Find the overlapping stables in L1 for each 32 stables. If none of L1 stable is
in compacting set then add them to compactionList. Also add them all to L0selected.
            
     
  }
  
   return compactionList;
}

Now compact each set in the list in parallel. 

                
> Make Level Compaction go faster(multiple independant compactions in parallel) for insert
heavy workload
> -------------------------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-4310
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-4310
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Core
>    Affects Versions: 1.1.1
>            Reporter: sankalp kohli
>            Priority: Minor
>              Labels: compaction, leveled
>
> Problem: If you are inserting data into cassandra and level compaction cannot catchup,
you will create lot of files in L0.  
> Here is a solution which will help here and also increase the performance of level compaction.
> We can do many compactions in parallel for unrelated data.
> 1) For no over lapping levels. Ex: when L0 stable is compacting with L1, we can do compactions
in other levels like L2 and L3 if they are eligible.
> 2) We can also do compactions with files in L1 which are not participating in L0 compactions.
> This is specially useful if you are using SSD and is not bottlenecked by IO. 
> I am seeing this issue in my cluster. The compactions pending are more than 50k.
> I am doing multithreaded to true and also not throttling the IO by putting the value
as 0. 
>  

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

Mime
View raw message