I can come up with a sample implementation and provide the result comparison of the performance measurement.
Daniel On 5/30/07, Marc Prud'hommeaux <[EMAIL PROTECTED]> wrote:
I personally think it sounds like a good idea that has a lot of potential for performance improvement. Perhaps someone could come up with a sample implementation that adds the API and a default implementation in the DataCacheImpl and compare the performance in the scenario mentioned below? That would help establish a concrete justification for enhancing the DataCache interface. On May 30, 2007, at 12:48 PM, Kevin Sutter wrote: > Marc, > What are your views on this request? Since you seem to be > intimately familiar with the data cache API, do you see a problem > with introducing this additional get method? Either from an > expectation viewpoint or an implementation viewpoint? Thanks. > > Kevin > > On 5/29/07, Daniel Lee <[EMAIL PROTECTED]> wrote: Hi Craig, > > The discussed API (getAll) is for fetching objects that's already > cached in > the DataCache. From what I understand, OpenJPA executes the > following code > when loading (find()) a customer which exists in the DataCache. It > loads > not only the customer but also the objects in any eager (direct and > indirect) relationships. In the earlier example (a customer with > 100 orders > and each order has different products) the direct relationships > are all orders placed by the customer and the indirect > relationships are all > products in these orders). > > 1. BrokerImpl.find() calls DataCacheStoreManager.initialize() to > initialize a new state manager for an object (a customer with > 100 orders > for example). > 2. initialize() then issues get() to DataCache to see whether > the data > (customer) is already cached. After successfully getting the > customer (data > != null) from the datacache, DataCachePCData.load (sm, fetch, > edata) is > invoked to load all the eager relationships (orders in the > example) of the > object (customer). > 3. PCDataImpl.load() loops through the relationship field to call > loadField() for each relationship which is not yet loaded. In > this example, > it is the relationship the customer to its orders (eager, > one-to-many) relationship > 4. loadField() calls toField() which is defined in AbstractPCData. > 5. toField() LOOPS through all elements (orders) to invoke > toNestedField() for each element. This is 100 toNestedFields > calls for the > 100 orders in the example. > 6. toNestedField() calls toRelationField(sm, vmd, data, fetch, > context) which actually calls find() and recursively get back to > step 1 > above for loading "a" order. This will end up calling get() 100 > times to > the DataCache for the 100 orders and can possibly get into > another loop for > loading all products in each order, etc. > > Because of the loop in step 5 above, a single "find(customerA)" > statement > actually triggers 100 DataCahce.get() for its orders and could be > hundreds > or thousands more of the get() calls for the products ordered by the > customer. This is a performance hit as I understand. > > If we have getAll(List keys) method which returns a list of objects > from the > datacache, we can change the logic to call the following new > methods to get > all elements (orders/products) in one relationship in single call to > getAll(); instead of calling get() a hundred times for 100 orders. > > - toNestedFields() - called by toFields without the loop > - toRelationFields() - called by toNestedFields; calls findAll() > - findAll need to be able to initialize a List of sm and call > initializeAll() > - initializeAll() - call getAll() instead of get(), then iterate > the > return to call load > > This is more like doing batch fetch from DataCache. There should > be some > significant performance improvement, especially in the distributed > environment in which the communication/serialization area is known > be the > bottleneck of the whole process. This implementation can also > potentially > provide a lot better performance for the 3-rd party DataCache plug- > ins which > provide and optimize getAll() process. > > Hope this make the issue more clear this time. Could you please > let me know > if you have further questions or other concerns. Many thanks. > > Daniel > > On 5/24/07, Craig L Russell <[EMAIL PROTECTED]> wrote: > > > Hi Daniel, > > > > On May 24, 2007, at 11:59 AM, Daniel Lee wrote: > > > > > Hi Craig, > > > > > > I think findAll() is different. It is a client level API and the > > > getAll() > > > here is for internal fetch from data cache. > > > > > > In the example, when an application issue findAll() for a list of > > > customers. It internally, for each customer with order(s), > loads the > > > "eager" relationship (orders) from data cache if they are already > > > cached by > > > calling map.get (orderId) for each order placed by the > customer. It > > > again > > > load the items that are related to each order by calling map.get > > > (itemId) for > > > each item if the relationship to Order is declared as eager. > This is > > > potentially a performance bottleneck and findAll() does not avoid > > > this. > > > > Seems that this algorithm can be improved to use the broker's > findAll > > mechanism when the instance is not found in the cache. The not-found > > instances can be found more efficiently than the code currently > does. > > > > Craig > > > > > > Thanks. > > > Daniel > > > > > > > > > On 5/23/07, Craig L Russell < [EMAIL PROTECTED]> wrote: > > >> > > >> Hi Daniel, > > >> > > >> Take a look at the findAll(Collection oids) method of > > >> OpenJPAEntityManager. This should do a better job than N get > (Object > > >> key) methods. > > >> > > >> Craig > > >> > > >> On May 23, 2007, at 3:55 PM, Daniel Lee wrote: > > >> > > >> > Do we miss the getAll(List keys) method for data cache? > > >> > > > >> > When fetching objects with eager "to-many" relationships, the > > >> code is > > >> > calling get(Object key) multiple time (one for each object > in the > > >> > relationship). For example, it is doing 1 get() call for each > > >> > order placed > > >> > by a customer which we are fetching, that means 100 calls for a > > >> > customer > > >> > with 100 orders. The performance can be greatly improved if we > > >> have > > >> > getAll(List keys) methods which returns all orders in one call. > > >> > This is > > >> > especially important in a distributed environment. > > >> > > > >> > Is there a way (new plug-in) to avoid the multiple-trip for > single > > >> > relationship, or can we implement the code to improve the > > >> > performance in > > >> > this area? > > >> > > > >> > Many thanks. > > >> > Daniel > > >> > > >> Craig Russell > > >> Architect, Sun Java Enterprise System http://java.sun.com/ > products/ > > >> jdo > > >> 408 276-5638 mailto:[EMAIL PROTECTED] > > >> P.S. A good JDO? O, Gasp! > > >> > > >> > > >> > > > > Craig Russell > > Architect, Sun Java Enterprise System http://java.sun.com/ > products/jdo > > 408 276-5638 mailto: [EMAIL PROTECTED] > > P.S. A good JDO? O, Gasp! > > > > > > > -- Marc Prud'hommeaux BEA Systems, Inc. Notice: This email message, together with any attachments, may contain information of BEA Systems, Inc., its subsidiaries and affiliated entities, that may be confidential, proprietary, copyrighted and/or legally privileged, and is intended solely for the use of the individual or entity named in this message. If you are not the intended recipient, and have received this message in error, please immediately return this by email and then delete it.
