[Mondrian] Re: Optimizing Sync Blocks in Aggregation class

Mon Jun 11 11:13:10 EDT 2007

Julian,

Thanks for the review.

I have checked in the Approach 2.

>>Re. our previous discussions on JDK 1.4 vs. 1.5. I understand that 
CopyOnWriteArrayList is only available under 1.5. Can we use a factory 
method to control the creation of this list, so that it can be made 
available in JDK 1.4 using retroweaver?

Retroweaver supports CopyOnWriteArrayList, we have test it. (Our previous 
discussions was about ReentrantReadWriteLock)

>>I don't really understand the profiler results. They seem to be improved 
in general, but there is a large number of blocks for HashMap in approach 
1. What caused that?

I didn't spend time on optimizing/analysing the approach 1 because we were 
inclined towards approach 2. 

>>What is the high-level locking strategy? In particular, if multiple 
objects need to be locked (say Aggregation and Segment), is there a 
well-defined locking order.
We have removed all locking on Aggregation(Segment list is synchronized by 
itself),  we are locking only on Segment and all the locking is done by 
method level synchronization, so we don't need to bother about explicit 
locking before any call. 

>> please document the locking strategy in the code
Updated details about fields and methods which are newly synchronized.

- Thiyagu

"Julian Hyde" <julianhyde at speakeasy.net> 
Sent by: mondrian-bounces at pentaho.org
11/06/2007 13:15
Please respond to
Mondrian developer mailing list <mondrian at pentaho.org>

To
"'Mondrian developer mailing list'" <mondrian at pentaho.org>
cc

Subject
RE: [Mondrian] Re: Optimizing Sync Blocks in Aggregation class

Thiyagu,

Both approaches seem to be an improvement. Approach 2 seems to be better, 
especially if the only mutable state in Aggregation is the list of 
Segments.

Re. our previous discussions on JDK 1.4 vs. 1.5. I understand that 
CopyOnWriteArrayList is only available under 1.5. Can we use a factory 
method to control the creation of this list, so that it can be made 
available in JDK 1.4 using retroweaver? 

I don't really understand the profiler results. They seem to be improved 
in general, but there is a large number of blocks for HashMap in approach 
1. What caused that?

What is the high-level locking strategy? In particular, if multiple 
objects need to be locked (say Aggregation and Segment), is there a 
well-defined locking order.

When you implement this, please document the locking strategy in the code. 
Maybe describe which fields are mutable and when, which classes to lock 
and in what order. Less is more - people are more likely to read, obey and 
maintain documentation which is terse and to the point.

I want to rationalize thread-locals out of the design and into real 
objects with a stated purpose and strategy. If you can start to do that 
with the design, I would appreciate it.

Lastly, please rename ConcurrentMDXTest to ConcurrentMdxTest; our coding 
standards call for acronyms to be InitCapped not CAPITALIZED.

Julian

PS Sorry I didn't reply earlier. This reply somehow ended up in my Drafts 
folder and I forgot to send it.

From: mondrian-bounces at pentaho.org [mailto:mondrian-bounces at pentaho.org] 
On Behalf Of Thiyagu Palanisamy
Sent: Thursday, May 31, 2007 5:42 AM
To: mondrian at pentaho.org
Subject: [Mondrian] Re: Optimizing Sync Blocks in Aggregation class

We have done some more improvements to our Approach 2 : Move the lock from 
Aggregation to Segment wherever it is possible (CR 9366) 

New Changes: 
1. Synchronized block removed from AggregationManager.getCellFromCache() 
The block was covering call to Aggregation.getCellValue() and 
synchronization is handled with in Aggregation.getCellValue() method 
itself. 
2. Synchronized block removed from Aggregation.getCellValue() 
This method was earlier synchronized to ensure that access to segmentRefs
is thread safe. Now we have made segmentRefs as CopyOnWriteArrayList 
instance instead of ArrayList. This frees us up from handling the 
Synchronized access to this list. We have selected CopyOnWriteArrayList 
because <quote java doc> it is more efficient than alternatives when 
traversal operations vastly outnumber mutations</quote java doc>, which is 
the case in our scenario, there will be lot of get cellvalue calls but 
only few times we will be modifying segments in the list. 
3. Made Segment.getCellValue() Synchronized 
        As we have removed synchronization from Aggregation we are 
synchronizing at Segment level. 
We have uploaded the latest change to 
http://forums.pentaho.org/showthread.php?t=54274 
We also ran Jprofiler against all the 3 versions(Current implementation, 
Approach1, Approach2) and the reports are attached to the same post. We 
could see that lock contention is drastically reduced for data load path 
(Aggregation object). 
Thanks, 
Thiyagu & Tushar 

Thiyagu Palanisamy/India/ThoughtWorks 
30/05/2007 20:29 

To
Mondrian 
cc

Subject
Optimizing Sync Blocks in Aggregation class

Hi, 

We are proposing optimization of synchronization blocks involved in 
loading of data from DB. 

Current system is synchronized on Aggregation object while their segments 
are loaded from DB. We did analysis using JProfiler to identify the 
maximum lock contentions. 

Issues with the current implementation: 
1. Because of the lock at Aggregation, requests are forced to wait even 
though they are not needed. Eg., If any previous request is loading a new 
segment to Aggregation all the requests that need the same aggregation 
will be blocked regardless of segment data availability. 
2. Aggregation is locked for entire period of creating segment, firing 
sql, loading result set back on to segments. So aggregation is locked when 
DB is processing data. 

We got 2 different working solutions to address this issue 

Approach 1: Keep locking on Aggregation itself, but optimization the 
duration lock (CR 9325) 
Changes made for this solution: 
1. Syncronized block removed from AggregationManager.load() 
    Because this blocks covers call to Aggregation.optimizePredicates(), 
Aggregation.load() but both these methods are already Synchronized. 
2. Made Aggregation.optimizePredicates() non synchronized
   Uses maxConstraints,star fields of aggregation instance but they are 
read only and set only in constructor. 
   Parameters to this method 
         columns : created at CellRequest, it is read only access in the 
method. This is related Aggregation object but columns list will remain 
same for the life time of the Aggregation class. 
         predicates : created at FBCR which is unique to a single request, 
it is also read only and has no association to Aggregation object. 
3. Made Aggregation.load() non synchronized
   Uses constrainedColumnsBitKey,segmentRefs fields of aggregation 
instance. 
          constrainedColumnsBitKey : This is read only and set in 
constructor 
          segmentRefs : We are still synchronizing on aggregation when 
accessing this 
     Parameters to this method 
          columns : created at CellRequest, this is set to columns 
instance variable. columns list will remain same for the life time of the 
Aggregation class so we have made change to set only at the first time 
load is called on aggregation. This will happen only during creation of 
Aggregation instance and it will be threadlocal at that point. 
          predicates : has no association to Aggregation object 
          pinnedSegments : has no association to Aggregation object 
          Measures: has no association to Aggregation object
4. Changes in Aggregation.load() 
   New Segments are added to segmentRefs only after they are loaded with 
the data. This will ensure that none of the Segments are exposed to other 
threads when they are loading. 
Advantages with this approach
   1. Aggregation object is locked only during the time of adding loaded 
Segments to segmentRefs.
   2. This approach requires only minor modifications to the existing 
locking strategy 
Disadvantage with this approach
   If more than one request comes at the same time for a same new segment 
of an existing aggregation, both the requests will fire query and load the 
segment data. This issue is already exists in the current version too, but 
window in which this can happen is small, this will happen only until one 
of the request reaches the AggregationManager.load(). With the new change 
this will happen until segment is set to AggregationManager.segmentRefs. 
Approach 2: Move the lock from Aggregation to Segment wherever it is 
possible (CR 9366) 
   1. Synchronized block removed from AggregationManager.load() : reason 
same as approach 1
   2. Made Aggregation.optimizePredicates() non synchronized : reason same 
as Approach 1
   3. Made Aggregation.load() non synchronized : reason same as Approach 1
   4. Changes in Aggregation.load() 
       Aggregation object is altered only when New Segments are added to 
segmentRefs, so we have Sync block covering add to segmentRefs     
   5. Made Aggregation.getCellValue() non synchronized 
       Only access to segmentRefs needs to be Synchronized, introduced 
sync block to cover access to segmentRefs and when segment is not ready 
and segment would contain the requested value(which means another thread 
is loading the segment) current request will be blocked until segment is 
loaded. And return value once segment is ready. If segment fails to load 
it will return null similar to when segment is not found. 
Advantage with this approach
   1.Aggregation object is locked only during the time of adding Segments 
to segmentRefs. 
   2. Concurrent requests will be blocked only if both the requests are 
for same segment which is loading. This is a significant improvement from 
Approach 1, because this reduces number of blocking requests. 
Disadvantage with this approach
   1.Locking strategy has been changed to part on Aggregation and part on 
Segment. This also means that we need more testing.
   2.We are using segment state for synchronization, this piece of code 
already exists but not used currently. 
We have attached both our implementations to the forum post: 
http://forums.pentaho.org/showthread.php?t=54274
ConcurrentMDXTest class is to detected lock issues, this test is skewed 
towards simulating concurrent access of same aggregation, this is not 
valid for performance comparison of different implementations. 

Please give your inputs on which approach we should move forward and 
issues or improvements we can make on to that approach. Also it would be 
of great help if you can give us some more scenarios for concurrent 
testing. 
Thanks,
Thiyagu, Tushar _______________________________________________
Mondrian mailing list
Mondrian at pentaho.org
http://lists.pentaho.org/mailman/listinfo/mondrian

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.pentaho.org/pipermail/mondrian/attachments/20070611/51fdea01/attachment-0001.html