This change is undesirable as it optimizes one path and makes several others behave in unintended ways. What happens if a vector with nulls shows up? What happens if a user sets a position to a null value in user code when this flag set?
If the answer to the above questions is the use is an advanced user, then why can't they just call: PlatformDependent.getInt(vector.memoryAddress() + position * 4). Why would we introduce something directly in the vector class for this specialized use case? If the user is advanced, that short memory access invocation seems fine to use. The whole idea with Arrow is that if you have a specialized algorithm, you can hand write memory reads and writes because you have predictable memory layout. On Wed, Sep 9, 2020 at 1:11 AM Fan Liya <liya.fa...@gmail.com> wrote: > Hi all, > > Thanks a lot for your previous feedback. > > Now we have made some investigation and prepared an initial PR supporting > the non-nullable IntVector [1], as this represents a common scenario. > Some initial observations and conclusions can be made. > > The basic idea of the PR is to provide a global static final flag > (NON_NULLABLE_VECTORS_ENABLED) to enable/disable the new feature. > 1. When the flag is enabled, manipulations to the validity buffer would be > by-passed if the vector is non-nullable. > 2. When the flag is disabled, the behavior is identical to the original > code. > > *First*, we want to show that the change is small. To support the > non-nullable IntVector, we need to change classes IntVector, > BaseFixedWidthVector (the superclass of IntVector), and BaseValueVector > (the superclass of BaseFixedWidthVector). The amount of changes (lines of > code) to each class is given below > > Class > > Additions (#lines) > > Deletions (#lines) > > Total # lines > > IntVector > > 12 > > 5 > > 370 > > BaseFixedWidthVector > > 30 > > 11 > > 925 > > BaseValueVector > > 7 > > 0 > > 240 > > It can be seen that the change is small, relative to the class size. In > addition, to support additional vector types, we only need to change the > sub-classes, and no more need to change the super classes. > > *Second*, we want to show that the performance improvement is notable. To > see this, we give the performance data of the IntBenchmark (with some > benchmarks added in the PR). To make the performance as good as possible, > we enable the ARROW_ENABLE_UNSAFE_MEMORY_ACCESS flag and disable > the ARROW_ENABLE_NULL_CHECK_FOR_GET flag. Below, we give the data with the > non-nullable vector flag turned off and on, respectively. > > > > > > > > > *(Vector non-nullable flag off)Benchmark Mode Cnt > Score Error UnitsIntBenchmarks.getInt avgt 5 > 384.948 ± 3.336 ns/opIntBenchmarks.isNull avgt 5 > 1301.005 ± 54.239 ns/opIntBenchmarks.setIntDirectly avgt 5 > 15387.486 ± 555.749 ns/opIntBenchmarks.setWithValueHolder avgt 5 > 15251.351 ± 134.286 ns/opIntBenchmarks.setWithWriter avgt 5 > 28595.586 ± 932.528 ns/op* > > *(Vector non-nullable flag on)* > > > > > *IntBenchmarks.getInt avgt 5 384.630 ± 1.547 > ns/opIntBenchmarks.isNull avgt 5 3.004 ± 0.110 > ns/opIntBenchmarks.setIntDirectly avgt 5 13511.605 ± 135.455 > ns/opIntBenchmarks.setWithValueHolder avgt 5 13035.883 ± 196.081 > ns/opIntBenchmarks.setWithWriter avgt 5 24734.825 ± 1603.708 > ns/op* > > For the *getInt *operation, there is little performance difference. This > is because we disable the ARROW_ENABLE_NULL_CHECK_FOR_GET flag, so > manipulations to the validity buffer are by-passed, even if the > non-nullable vector flag is off. When the ARROW_ENABLE_NULL_CHECK_FOR_GET > is enabled, there is a 72.4% performance improvement gained by turning on > the non-nullable vector flag. > > For the isNull operation, we see 3 orders of magnitude performance > improvements by enabling the non-nullable vector flag. > > For other operations, we see 12.2%, 13.5% and 14.5% performance > improvements by turning on the non-nullable vector flag. > > So it can be seen that notable performance improvements can be gained for > non-nullable vectors. > > *Third*, we want to claim that for nullable vectors and scenarios when we > turn off the non-nullable vector flag, the new changes do not introduce > performance regression. Such concern is plausible, as our changes add some > if-else branches to the code, which may degrade performance. > > We give the benchmark results of the original code, as below. > > *(original ocde)* > > > > > > *Benchmark Mode Cnt Score Error > UnitsIntBenchmarks.getInt avgt 5 383.511 ± 0.156 > ns/opIntBenchmarks.isNull avgt 5 1274.271 ± 19.092 > ns/opIntBenchmarks.setIntDirectly avgt 5 15162.219 ± 194.956 > ns/opIntBenchmarks.setWithValueHolder avgt 5 15247.640 ± 153.103 > ns/opIntBenchmarks.setWithWriter avgt 5 28587.780 ± 160.458 > ns/op* > > By comparing this set of results with the above results with non-nullable > vector flag disabled, little performance difference can be observed, > indicating no performance regression. > > By examining the generated assembly, it can be seen that JIT is smart > enough to remove the if-else branch completely (the below screen-shot gives > an example where the if branch is optimized away in the assembly) > > So if some users do not like this feature, because their vectors are > always nullable, they can simply disable the flag, and no performance > difference can be observed. > > Would you please give some feedback? > If it looks good to you, maybe we can go ahead to support other types of > vectors. > Thank you in advance. > > Liya Fan > > [image: image.png] > > [1] https://github.com/apache/arrow/pull/8147 > > On Fri, Mar 13, 2020 at 9:47 PM Fan Liya <liya.fa...@gmail.com> wrote: > >> Hi Jacques, >> >> Thanks a lot for your valuable comments. >> >> I agree with you that collapsing nullable and non-nullable >> implementations is a good idea, and it does not contradict with the idea of >> introducing a fast code path, if it does not introduce much cost or code >> complexity. >> >> The idea of word level checking is interesting. >> >> As you suggested, we will do more investigations and reconsider the >> problem from a broader perspective. >> >> Best, >> Liya Fan >> >> >> >> On Thu, Mar 12, 2020 at 9:27 AM Jacques Nadeau <jacq...@apache.org> >> wrote: >> >>> Generally Ive found that this isnt an important optimization in the use >>> cases we see. Memory overhead, especially with our Java shared allocation >>> scheme is nominal. Optimizing null checks at the word level usually is >>> much >>> more impactful since non null and null runs are much more common on a >>> shorter window common than they seeing those declared. >>> >>> In other words, I'd suggest you look at your problem with a broader >>> perspective and see whether you're actually focused on optimizing the >>> most >>> important dimension. >>> >>> As an aside, the original Arrow Java code actually treated these concepts >>> more distinctly and we consciously made a decision to collapse them to >>> simplify real world use. >>> >>> I do think it makes to add a dirty read interface if you want. This would >>> allow consumers of the interface to behave efficiently if they wanted to. >>> >>> One last note, efficient evaluation and processing should generally >>> always >>> work at the validity word level. Adding an extra if check at the word >>> versus extra complexity of having an early out per batch seems like a >>> pretty small life in the grand scheme of processing. >>> >>> On Wed, Mar 11, 2020, 9:15 AM Brian Hulette <hulet...@gmail.com> wrote: >>> >>> > > And there is a "nullable" metadata-only flag at the >>> > > Field level. Could the same kinds of optimizations be implemented in >>> > > Java without introducing a "nullable" concept? >>> > >>> > Note Liya Fan did suggest pulling the nullable flag from the Field >>> when the >>> > vector is created in item (1) of the proposed changes. >>> > >>> > Brian >>> > >>> > On Wed, Mar 11, 2020 at 5:54 AM Fan Liya <liya.fa...@gmail.com> wrote: >>> > >>> > > Hi Micah, >>> > > >>> > > Thanks a lot for your valuable comments. Please see my comments >>> inline. >>> > > >>> > > > I'm a little concerned that this will change assumptions for at >>> least >>> > > some >>> > > > of the clients using the library (some might always rely on the >>> > validity >>> > > > buffer being present). >>> > > >>> > > I can understand your concern and I am also concerned. >>> > > IMO, the client should not depend on this assumption, as the >>> > specification >>> > > says "Arrays having a 0 null count may choose to not allocate the >>> > validity >>> > > bitmap." [1] >>> > > That being said, I think it would be safe to provide a global flag to >>> > > switch on/off the feature (as you suggested). >>> > > >>> > > > I think this is a good feature to have for the reasons you >>> mentioned. >>> > It >>> > > > seems like there would need to be some sort of configuration bit >>> to set >>> > > for >>> > > > this behavior. >>> > > >>> > > Good suggestion. We should be able to switch on and off the feature >>> with >>> > a >>> > > single global flag. >>> > > >>> > > > But, I'd be worried about code complexity this would >>> > > > introduce. >>> > > >>> > > I agree with you that code complexity is an important factor to >>> consider. >>> > > IMO, our proposal should not involve too much code change, or >>> increase >>> > code >>> > > complexity too much. >>> > > To prove this, maybe we need to show some small experimental code >>> change. >>> > > >>> > > Best, >>> > > Liya Fan >>> > > >>> > > [1] https://arrow.apache.org/docs/format/Columnar.html#logical-types >>> > > >>> > > On Wed, Mar 11, 2020 at 1:53 PM Micah Kornfield < >>> emkornfi...@gmail.com> >>> > > wrote: >>> > > >>> > > > Hi Liya Fan, >>> > > > I'm a little concerned that this will change assumptions for at >>> least >>> > > some >>> > > > of the clients using the library (some might always rely on the >>> > validity >>> > > > buffer being present). >>> > > > >>> > > > I think this is a good feature to have for the reasons you >>> mentioned. >>> > It >>> > > > seems like there would need to be some sort of configuration bit >>> to set >>> > > for >>> > > > this behavior. But, I'd be worried about code complexity this would >>> > > > introduce. >>> > > > >>> > > > Thanks, >>> > > > Micah >>> > > > >>> > > > On Tue, Mar 10, 2020 at 6:42 AM Fan Liya <liya.fa...@gmail.com> >>> wrote: >>> > > > >>> > > > > Hi Wes, >>> > > > > >>> > > > > Thanks a lot for your quick reply. >>> > > > > I think what you mentioned is almost exactly what we want to do >>> in >>> > > > Java.The >>> > > > > concept is not important. >>> > > > > >>> > > > > Maybe there are only some minor differences: >>> > > > > 1. In C++, the null_count is mutable, while for Java, once a >>> vector >>> > is >>> > > > > constructed as non-nullable, its null count can only be 0. >>> > > > > 2. In C++, a non-nullable array's validity buffer is null, while >>> in >>> > > Java, >>> > > > > the buffer is an empty buffer, and cannot be changed. >>> > > > > >>> > > > > Best, >>> > > > > Liya Fan >>> > > > > >>> > > > > On Tue, Mar 10, 2020 at 9:26 PM Wes McKinney < >>> wesmck...@gmail.com> >>> > > > wrote: >>> > > > > >>> > > > > > hi Liya, >>> > > > > > >>> > > > > > In C++ we elect certain faster code paths when the null count >>> is 0 >>> > or >>> > > > > > computed to be zero. When the null count is 0, we do not >>> allocate a >>> > > > > > validity bitmap. And there is a "nullable" metadata-only flag >>> at >>> > the >>> > > > > > Field level. Could the same kinds of optimizations be >>> implemented >>> > in >>> > > > > > Java without introducing a "nullable" concept? >>> > > > > > >>> > > > > > - Wes >>> > > > > > >>> > > > > > On Tue, Mar 10, 2020 at 8:13 AM Fan Liya <liya.fa...@gmail.com >>> > >>> > > wrote: >>> > > > > > > >>> > > > > > > Dear all, >>> > > > > > > >>> > > > > > > A non-nullable vector is one that is guaranteed to contain no >>> > > nulls. >>> > > > We >>> > > > > > > want to support non-nullable vectors in Java. >>> > > > > > > >>> > > > > > > *Motivations:* >>> > > > > > > 1. It is widely used in practice. For example, in a database >>> > > engine, >>> > > > a >>> > > > > > > column can be declared as not null, so it cannot contain null >>> > > values. >>> > > > > > > 2.Non-nullable vectors has significant performance advantages >>> > > > compared >>> > > > > > with >>> > > > > > > their nullable conterparts, such as: >>> > > > > > > 1) the memory space of the validity buffer can be saved. >>> > > > > > > 2) manipulation of the validity buffer can be bypassed >>> > > > > > > 3) some if-else branches can be replaced by sequential >>> > > instructions >>> > > > > (by >>> > > > > > > the JIT compiler), leading to high throughput for the CPU >>> > pipeline. >>> > > > > > > >>> > > > > > > *Potential Cost:* >>> > > > > > > For nullable vectors, there can be extra checks against the >>> > > > > nullablility >>> > > > > > > flag. So we must change the code in a way that minimizes the >>> > cost. >>> > > > > > > >>> > > > > > > *Proposed Changes:* >>> > > > > > > 1. There is no need to create new vector classes. We add a >>> final >>> > > > > boolean >>> > > > > > to >>> > > > > > > the vector base classes as the nullability flag. The value >>> of the >>> > > > flag >>> > > > > > can >>> > > > > > > be obtained from the field when creating the vector. >>> > > > > > > 2. Add a method "boolean isNullable()" to the root interface >>> > > > > ValueVector. >>> > > > > > > 3. If a vector is non-nullable, its validity buffer should >>> be an >>> > > > empty >>> > > > > > > buffer (not null, so much of the existing logic can be left >>> > > > unchanged). >>> > > > > > > 4. For operations involving validity buffers (e.g. isNull, >>> get, >>> > > set), >>> > > > > we >>> > > > > > > use the nullability flag to bypass manipulations to the >>> validity >>> > > > > buffer. >>> > > > > > > >>> > > > > > > Therefore, it should be possible to support the feature with >>> > small >>> > > > code >>> > > > > > > changes. >>> > > > > > > >>> > > > > > > BTW, please note that similar behaviors have already been >>> > supported >>> > > > in >>> > > > > > C++. >>> > > > > > > >>> > > > > > > Would you please give your valueable feedback? >>> > > > > > > >>> > > > > > > Best, >>> > > > > > > Liya Fan >>> > > > > > >>> > > > > >>> > > > >>> > > >>> > >>> >>
