[elephant-devel] working with many millions of objects

Ian Eslick eslick at csail.mit.edu
Fri Feb 16 06:55:20 UTC 2007


Some experiments performed on an old topic...

On Oct 11, 2006, at 8:57 PM, Red Daly wrote:

> I was importing into sleepycat using standard elephant routines.  I  
> am not aware of an 'import mode' for sleepycat, but I will look  
> into that when I have a chance.  Another consideration using  
> sleepycat is that using BTrees with a large working set demands  
> large amounts of memory  relative to a Hash representation.  I am  
> unfamiliar with the internals of elephant and sleepycat, but it  
> feels like the basic access method is restricting performance,  
> which seems to be described here:
> http://www.sleepycat.com/docs/gsg/C/accessmethods.html#BTreeVSHash
>
> My problem so far has been importing the data, which goes very fast  
> until sleepycat requires extensive disk access.  The in-memory rate  
> is reasonable and would complete in a few hours.  However, once  
> disk operations begin the import speed suggests it would take many  
> days to complete.  I have yet to perform extensive benchmarks, but  
> I estimate the instantiation rate shifts from 1800 persistent class  
> instantiations /second to 120 / s.

The biggest performance factor is properly managing transaction sizes  
to balance contention, total locks used, etc.

You can also turn off all transactional and log synchronization since  
if you crash, you can always restart a several hour download.  I  
think this may avoid additional overhead, however I have not  
benchmarked this.  i.e.

(with-transaction (:txn-nosync t :dirty-read t)
    (create 500 objects))


> here are the two approaches that I hypothesize may help  
> performance.  I am admittedly unaware of innards of the two systems  
> in question, so you expert developers will know best.  If either  
> sounds appropriate or you envision another possibility for allowing  
> this kind of scaling, I will look into implementing such a system.
>
> 1.  decreasing the size of the working set is one possibility for  
> decreasing run-time memory requirements and disk access.  I'm not  
> sure how the concept of a 'working set' translates from the  
> sleepycat world to the elephant world, but perhaps you do.

What do you mean by working set?  When loading stuff into a database  
you are moving index pages around in the btree and allocating endless  
amounts of leaf nodes.  The index nodes are cacheable, but the leaf  
nodes are definitely not!  I think there are ways to add a bunch of  
objects and then force a btree to update all the index pages.    
Access to that functionality, if BDB even supports it, is not  
provided in elephant.

> 2.  using a Hash instead of a BTree in the primary database?  I am  
> unsure what this means for elephant.

I finally got around to trying this and it showed poorer performance  
on a large stress test (create, modify and access 10k objects).  I  
don't have a good theory as to why it was slower other than in create  
where the hash table had to grow.

>
> In the mean time I will depart from the every-class-is-persistent  
> approach and also use more traditional data structures.
>
> Thanks again,
> Red Daly
>
>
>
> Robert L. Read wrote:
>> Yes, it's amusing.
>>
>> In my own work I use the Postgres backend; I know very little  
>> about SleepyCat.  It seems
>> to me that this is more of a SleepyCat issue, then an Elephant  
>> issue.  Perhaps you should
>> ask the SleepyCat list?
>>
>> Are you importing things into SleepyCat directly in the correct  
>> serialization format that
>> they can be read by Elephant?  If so, I assume it is just a  
>> question of solving the SleepyCat
>> problems.
>>
>> An alternative would be to use the SQL-based backend.  However, I  
>> doubt this will solve
>> your problem, since at present we (well, I wrote it) use a very  
>> inefficient serialization scheme
>> for the SQL-based backend that base64 encodes everything.  This  
>> had the advantage that
>> it makes it work trouble-free with different database backends,  
>> but could clearly be improved upon.
>> However, it is more than efficient enough for all my work, and at  
>> present nobody is clamoring
>> to have it improved.
>>
>> Is your problem importing the data or using it once it is  
>> imported?  It's hard for me to imagine
>> a problem so large that even the import time is a problem ---  
>> suppose it takes 24 hours --- can
>> you not afford to pay that?
>>
>> A drastic measure and potentially expensive measure would be to  
>> switch to a 64-bit architecture
>> with a huge memory.  I intend to do that when forced by  
>> performance issues in my own work.
>>
>>
>>
>> On Tue, 2006-10-10 at 00:46 -0700, Red Daly wrote:
>>> I will be running experiments in informatics and modeling in the  
>>> future that may contain (tens or hundreds of) millions of  
>>> objects.  Given the ease of use of elephant so far, it would be  
>>> great to use it as the persistent store and avoid creating too  
>>> many custom data structures.
>>>
>>> I have recently run up against some performance bottlenecks when  
>>> using elephant to work with very large datasets (in the hundreds  
>>> of millions of objects).  Using SleepyCat, I am able to import  
>>> data very quickly with a DB_CONFIG file with the following contents:
>>>
>>> set_lk_max_locks 500000
>>> set_lk_max_objects 500000
>>> set_lk_max_lockers 500000
>>> set_cachesize 1 0 0
>>>
>>> I can import data very quickly until the 1 gb cache is too small  
>>> to allow complete in-memory access to the database.  at this  
>>> point it seems that disk IO makes additional writes happen much  
>>> slower.  (I have also tried increasing the 1 gb cache size, but  
>>> the database fails to open if it is too large--e.g. 2 gbs.  I  
>>> have 1.25 gb physical memory and 4 gb swap, so the constraint  
>>> seems to be physical memory.)  the max_lock, etc. lines allow  
>>> transactions to contain hundreds of thousands of individual  
>>> locks, limiting the transaction throughput bottleneck
>>>
>>> What are the technical restrictions on writing several million  
>>> objects to the datastore?  Is it feasible to create a batch  
>>> import feature to allow large datasets to be imported using  
>>> reasonable amounts of memory for a desktop computer?
>>>
>>> I hope this email is at least amusing!
>>>
>>> Thanks again,
>>> red daly
>>> _______________________________________________
>>> elephant-devel site list
>>> elephant-devel at common-lisp.net <mailto:elephant-devel at common- 
>>> lisp.net>
>>> http://common-lisp.net/mailman/listinfo/elephant-devel
>>>
>> --------------------------------------------------------------------- 
>> ---
>>
>> _______________________________________________
>> elephant-devel site list
>> elephant-devel at common-lisp.net
>> http://common-lisp.net/mailman/listinfo/elephant-devel
>
> _______________________________________________
> elephant-devel site list
> elephant-devel at common-lisp.net
> http://common-lisp.net/mailman/listinfo/elephant-devel




More information about the elephant-devel mailing list