Sorry I could not assist the meeting live but reviewed the logs and had some remarks, so ehre are the logs and my remarks :) ## Meeting logs http://transcripts.jboss.org/meeting/irc.freenode.org/hibernate-dev/2012/... Minutes (text):http://transcripts.jboss.org/meeting/irc.freenode.org/hibernate-dev... Log: http://transcripts.jboss.org/meeting/irc.freenode.org/hibernate-dev/2012/... ## About parallel arrays vs arrays of objects Even today, there is a cost in using Objects or HashMap as data placeholders. This might not be relevant for us especially since we are not using them for "live" data but since memory footprint has always been a concern for Hibernate, it's worth mentioning. This presentation is quite interesting and show the extra memory cost of such structure compared to arrays http://www.cs.virginia.edu/kim/publicity/pldi09tutorials/memory-efficient... ## PessimisticLockException versus LockAcquisitionException I am not certain of that and Git history got cut short with the migration of entitymanager into core. But I do have recollections of some lock exceptions that were not catch up and I had to use PessimisticLockException. It could have been a change triggered by the TCK. Again I might be wrong, it was a long time ago. Plus ORM's core has evolved since. Emmanuel _______________________________________________ hibernate-dev mailing list hibernate-dev(a)lists.jboss.org https://lists.jboss.org/mailman/listinfo/hibernate-dev

Performance I don't know, you are probably right. But memory wise, that could be way different. Even ignoring the overhead of the object + pointer in memory, the alignment of boolean or other small objects would make a significant impact. Of course if we are talking about 20 values, we should not bother. But persisters and the like store more than 20 values and we have more than one persister / loader. It might be inconsequential in the end but that might be worth testing. On a related note it's up for debate whether or not putting data in a hash map for faster lookup later is worth it in all cases: - it takes much more space than raw arrays - array scan might be as fast or faster for a small enough array. As we have seen in Infinispan and OGM, computing a hash is not a cheap operation. Again this require testing but I am guilty as charge of using collections in AnnotationBinder when doing some computations that would be better off written as an array + array scan. On 3 mai 2012, at 19:32, Steve Ebersole wrote: > I seriously doubt the performance cost of 20 'parallel arrays' versus 1 array of Objects holding those 20 values is anything but negligible at best. > > > On Thu 03 May 2012 11:04:30 AM CDT, Emmanuel Bernard wrote: >> Sorry I could not assist the meeting live but reviewed the logs and had some remarks, so ehre are the logs and my remarks :) >> >> ## Meeting logs >> >> http://transcripts.jboss.org/meeting/irc.freenode.org/hibernate-dev/2012/... >> Minutes (text):http://transcripts.jboss.org/meeting/irc.freenode.org/hibernate-dev... >> Log:            http://transcripts.jboss.org/meeting/irc.freenode.org/hibernate-dev/2012/hibernate-dev.2012-05-03-14.01.log.html >> >> ## About parallel arrays vs arrays of objects >> >> Even today, there is a cost in using Objects or HashMap as data placeholders. This might not be relevant for us especially since we are not using them for "live" data but since memory footprint has always been a concern for Hibernate, it's worth mentioning. >> >> This presentation is quite interesting and show the extra memory cost of such structure compared to arrays http://www.cs.virginia.edu/kim/publicity/pldi09tutorials/memory-efficient... >> >> ## PessimisticLockException versus LockAcquisitionException >> >> I am not certain of that and Git history got cut short with the migration of entitymanager into core. But I do have recollections of some lock exceptions that were not catch up and I had to use PessimisticLockException. It could have been a change triggered by the TCK. Again I might be wrong, it was a long time ago. Plus ORM's core has evolved since. >> >> Emmanuel >> _______________________________________________ >> hibernate-dev mailing list >> hibernate-dev(a)lists.jboss.org >> https://lists.jboss.org/mailman/listinfo/hibernate-dev > > -- > steve(a)hibernate.org > http://hibernate.org _______________________________________________ hibernate-dev mailing list hibernate-dev(a)lists.jboss.org https://lists.jboss.org/mailman/listinfo/hibernate-dev

Even taking the risk of pouring oil onto the fire, I think a simpler data structure wins in most cases over the parallel arrays. It is much harder to use the latter and easier to make mistakes which leads to more bugs and higher maintenance costs. As Sanne is saying performance questions are tricky. So many thing are happening with the code our days before they are getting executed on the bare metal that it is hard to know what performance impacts a certain change has. In the end you just have to measure. Personally I think we should primarily strive for a better and easier to use API. Oppertunities to optimizes arise then often naturally. And now my dear disciples let me close with: "The First Rule of Program Optimization: Don't do it. The Second Rule of Program Optimization (for experts only!): Don't do it yet." — Michael A. Jackson :-) --Hardy On May 4, 2012, at 11:58 AM, Sanne Grinovero wrote: > tricky subject. > I'm confident that there are many cases in which we should have used > arrays rather than maps, especially for temporary objects which aren't > short lived enough (an HashMap living in the scope of a single method > is going to be cheap). We should have either objects allocated for > very long (like forever in the scope of the SessionFactory), or very > short. > > In the case of how we keep metadata, I think performance would be > dominated not that much by the fact it's a slightly bigger object but > by prefetching and what is going to be available in the cache lines > you just have filled in: obviously cache is way faster than memory so > being clever in the sequence you lay out your data structure could > speed you up by a couple of orders of magnitude. > > Using primitives and array matrixes makes the data smaller, hence more > likely to fit in the cache; but if using an array of objects in which > each object collects the needed fields in one group, that's likely > going to be faster.. but I'm making assumptions on how this structure > is going to be read more frequently. > > For example when declaring a matrix as an [ ][ ], performance will be > very different depending if you read by columns or rows - forgot which > one is better now - but in that case if the common use case is using > the slower path it's usually a good idea to invert the matrix. > > I'd love it if we could enter this space, or even if it's not suited > for it, at least be considered "lite": > http://stackoverflow.com/questions/10374735/lucene-and-ormlite-on-android > > Sanne > > On 4 May 2012 10:07, Emmanuel Bernard <emmanuel(a)hibernate.org&gt; wrote: >> Performance I don't know, you are probably right. But memory wise, that could be way different. >> Even ignoring the overhead of the object + pointer in memory, the alignment of boolean or other small objects would make a significant impact. >> >> Of course if we are talking about 20 values, we should not bother. But persisters and the like store more than 20 values and we have more than one persister / loader. It might be inconsequential in the end but that might be worth testing. >> >> On a related note it's up for debate whether or not putting data in a hash map for faster lookup later is worth it in all cases: >> >> - it takes much more space than raw arrays >> - array scan might be as fast or faster for a small enough array. As we have seen in Infinispan and OGM, computing a hash is not a cheap operation. >> >> Again this require testing but I am guilty as charge of using collections in AnnotationBinder when doing some computations that would be better off written as an array + array scan. >> >> >> On 3 mai 2012, at 19:32, Steve Ebersole wrote: >> >>> I seriously doubt the performance cost of 20 'parallel arrays' versus 1 array of Objects holding those 20 values is anything but negligible at best. _______________________________________________ hibernate-dev mailing list hibernate-dev(a)lists.jboss.org https://lists.jboss.org/mailman/listinfo/hibernate-dev

Yes, but we can't blindly go for the nicer approach: we are a library and hence used by many. Look at the bottleneck we found on Hibernate ORM due to the change of magnitude Hibernate OGM involved. Look at the perf issues we found in OGM itself just because I used the builder pattern for some objects using in critical paths. Same for hashCode that was not cached in some critical objects. I'm all fine with the nice approach if it's followed by a perf test before being pushed :) On 4 mai 2012, at 12:39, Hardy Ferentschik wrote: > Even taking the risk of pouring oil onto the fire, I think a simpler data structure wins in most cases over > the parallel arrays. It is much harder to use the latter and easier to make mistakes which leads to more > bugs and higher maintenance costs. > > As Sanne is saying performance questions are tricky. So many thing are happening with the code our days > before they are getting executed on the bare metal that it is hard to know what performance impacts a certain > change has. In the end you just have to measure. > > Personally I think we should primarily strive for a better and easier to use API. Oppertunities to optimizes arise > then often naturally. > > And now my dear disciples let me close with: > "The First Rule of Program Optimization: Don't do it. The Second Rule of Program Optimization (for experts only!): Don't do it yet." — Michael A. Jackson > > :-) > > --Hardy > > On May 4, 2012, at 11:58 AM, Sanne Grinovero wrote: > >> tricky subject. >> I'm confident that there are many cases in which we should have used >> arrays rather than maps, especially for temporary objects which aren't >> short lived enough (an HashMap living in the scope of a single method >> is going to be cheap). We should have either objects allocated for >> very long (like forever in the scope of the SessionFactory), or very >> short. >> >> In the case of how we keep metadata, I think performance would be >> dominated not that much by the fact it's a slightly bigger object but >> by prefetching and what is going to be available in the cache lines >> you just have filled in: obviously cache is way faster than memory so >> being clever in the sequence you lay out your data structure could >> speed you up by a couple of orders of magnitude. >> >> Using primitives and array matrixes makes the data smaller, hence more >> likely to fit in the cache; but if using an array of objects in which >> each object collects the needed fields in one group, that's likely >> going to be faster.. but I'm making assumptions on how this structure >> is going to be read more frequently. >> >> For example when declaring a matrix as an [ ][ ], performance will be >> very different depending if you read by columns or rows - forgot which >> one is better now - but in that case if the common use case is using >> the slower path it's usually a good idea to invert the matrix. >> >> I'd love it if we could enter this space, or even if it's not suited >> for it, at least be considered "lite": >> http://stackoverflow.com/questions/10374735/lucene-and-ormlite-on-android >> >> Sanne >> >> On 4 May 2012 10:07, Emmanuel Bernard&lt;emmanuel(a)hibernate.org&gt; wrote: >>> Performance I don't know, you are probably right. But memory wise, that could be way different. >>> Even ignoring the overhead of the object + pointer in memory, the alignment of boolean or other small objects would make a significant impact. >>> >>> Of course if we are talking about 20 values, we should not bother. But persisters and the like store more than 20 values and we have more than one persister / loader. It might be inconsequential in the end but that might be worth testing. >>> >>> On a related note it's up for debate whether or not putting data in a hash map for faster lookup later is worth it in all cases: >>> >>> - it takes much more space than raw arrays >>> - array scan might be as fast or faster for a small enough array. As we have seen in Infinispan and OGM, computing a hash is not a cheap operation. >>> >>> Again this require testing but I am guilty as charge of using collections in AnnotationBinder when doing some computations that would be better off written as an array + array scan. >>> >>> >>> On 3 mai 2012, at 19:32, Steve Ebersole wrote: >>> >>>> I seriously doubt the performance cost of 20 'parallel arrays' versus 1 array of Objects holding those 20 values is anything but negligible at best. > > > _______________________________________________ > hibernate-dev mailing list > hibernate-dev(a)lists.jboss.org > https://lists.jboss.org/mailman/listinfo/hibernate-dev _______________________________________________ hibernate-dev mailing list hibernate-dev(a)lists.jboss.org https://lists.jboss.org/mailman/listinfo/hibernate-dev

Apparently this did not go through to the list the first time, sorry... Completely agree. I focused on perf in my last comment but I dont think memory is all that much different. The declaration of all that state already has to be accounted for in its current "flattened" parallel-array representation. The trades off here are: 1) X number of array declarations versus 1 2) overhead of the class definition; again its actual state field memory footprint is already accounted for so we really are just talking about small amount of memory here. Certainly I think its a great idea to try to actually calculate and compare the memory diffs here. I am pretty confident the difference is negligible. But either way Hardy's point about higher likelihood of bugs is the biggest concern. In my experience lack of cohesive encapsulation is just a recipe for situations where hard to find problems creep into the code. On 05/04/2012 05:39 AM, Hardy Ferentschik wrote: > Even taking the risk of pouring oil onto the fire, I think a simpler data structure wins in most cases over > the parallel arrays. It is much harder to use the latter and easier to make mistakes which leads to more > bugs and higher maintenance costs. > > As Sanne is saying performance questions are tricky. So many thing are happening with the code our days > before they are getting executed on the bare metal that it is hard to know what performance impacts a certain > change has. In the end you just have to measure. > > Personally I think we should primarily strive for a better and easier to use API. Oppertunities to optimizes arise > then often naturally. > > And now my dear disciples let me close with: > "The First Rule of Program Optimization: Don't do it. The Second Rule of Program Optimization (for experts only!): Don't do it yet." — Michael A. Jackson > > :-) > > --Hardy > > On May 4, 2012, at 11:58 AM, Sanne Grinovero wrote: > >> tricky subject. >> I'm confident that there are many cases in which we should have used >> arrays rather than maps, especially for temporary objects which aren't >> short lived enough (an HashMap living in the scope of a single method >> is going to be cheap). We should have either objects allocated for >> very long (like forever in the scope of the SessionFactory), or very >> short. >> >> In the case of how we keep metadata, I think performance would be >> dominated not that much by the fact it's a slightly bigger object but >> by prefetching and what is going to be available in the cache lines >> you just have filled in: obviously cache is way faster than memory so >> being clever in the sequence you lay out your data structure could >> speed you up by a couple of orders of magnitude. >> >> Using primitives and array matrixes makes the data smaller, hence more >> likely to fit in the cache; but if using an array of objects in which >> each object collects the needed fields in one group, that's likely >> going to be faster.. but I'm making assumptions on how this structure >> is going to be read more frequently. >> >> For example when declaring a matrix as an [ ][ ], performance will be >> very different depending if you read by columns or rows - forgot which >> one is better now - but in that case if the common use case is using >> the slower path it's usually a good idea to invert the matrix. >> >> I'd love it if we could enter this space, or even if it's not suited >> for it, at least be considered "lite": >> http://stackoverflow.com/questions/10374735/lucene-and-ormlite-on-android >> >> Sanne >> >> On 4 May 2012 10:07, Emmanuel Bernard&lt;emmanuel(a)hibernate.org&gt; wrote: >>> Performance I don't know, you are probably right. But memory wise, that could be way different. >>> Even ignoring the overhead of the object + pointer in memory, the alignment of boolean or other small objects would make a significant impact. >>> >>> Of course if we are talking about 20 values, we should not bother. But persisters and the like store more than 20 values and we have more than one persister / loader. It might be inconsequential in the end but that might be worth testing. >>> >>> On a related note it's up for debate whether or not putting data in a hash map for faster lookup later is worth it in all cases: >>> >>> - it takes much more space than raw arrays >>> - array scan might be as fast or faster for a small enough array. As we have seen in Infinispan and OGM, computing a hash is not a cheap operation. >>> >>> Again this require testing but I am guilty as charge of using collections in AnnotationBinder when doing some computations that would be better off written as an array + array scan. >>> >>> >>> On 3 mai 2012, at 19:32, Steve Ebersole wrote: >>> >>>> I seriously doubt the performance cost of 20 'parallel arrays' versus 1 array of Objects holding those 20 values is anything but negligible at best. > > _______________________________________________ > hibernate-dev mailing list > hibernate-dev(a)lists.jboss.org > https://lists.jboss.org/mailman/listinfo/hibernate-dev