Columnarization in Rust

http://www.frankmcsherry.org/columnarization/serialization/rust/2014/12/15/Columnarization-in-Rust.html

by steveklabnik • 12 years ago

75 27 12 years ago

frankmcsherry.org

arthursilva• 12 years ago

Make sure you compile your release/bench builds with --release (if using cargo) or (-O if using rustc).

masklinn• 12 years ago

IIRC release is the default when testing or benching.

dbaupp• 12 years ago

Only when using `cargo bench` to compile/run them.

kbenson• 12 years ago

So like everyone without a job, I’ve started to learn Rust. And like everyone who has started to learn Rust, I now feel it is very important to tell you about my experience with it.

Well, that sums that up pretty succinctly. :)

rubiquity• 12 years ago

That's the best intro to a blog post that I've ever read.

frou_dh• 12 years ago

It gets two points deducted for the plague of the unnecessary introductory So.

thomasahle• 12 years ago

Seeing how the Rust community seems to have converged on using `uint` as the default type for array indexing and so on; wouldn't it make sense to reflect that in the complexity of the name of the types?

Instead of (uint, int) we could have (int, sint) or (uint, sint).

dilap• 12 years ago

You are going to confuse a hell of a lot of people if you make int mean "unsigned integer"!

long• 12 years ago

So maybe "nat" for "uint", then?

steveklabnik2• 12 years ago

wouldn't it make sense to reflect that in the complexity of the name of the types?

There has been a huuuuuge amount of discussion about int/uint and naming, expect some clarification sometime soonish, as any changes (if any) would have to be made soon.

Igglyboo• 12 years ago

No because it will confuse people from pretty much every other language that uses 'int' to mean 'signed integer'.

zcdziura• 12 years ago

Just for legacy sake, it probably wouldn't make sense to switch the meaning of 'int' in Rust to 'uint', whilst creating a new type 'sint' to emulate the old meaning of 'int'. Legions of up-and-coming developers have learned what an integer is, and it would be incredibly counter-productive to change up that definition for one particular programming language.

vowelless• 12 years ago

In addition to what the others have said, it would also be semantically incorrect. Integers in mathematics are signed. So that terminology would not cross over very well.

heinrich5991• 12 years ago

Yea, and it is being misused in a lot of places. In this project, `u32` would probably be a better fit as you try to serialize data – this should happen in a platform-independent way.

skybrian• 12 years ago

I'm not familiar with Rust, but I wonder how much copying this does for deeply nested types?

steveklabnik2• 12 years ago

Rust has move semantics by default. A move does a bytewise copy, and then disallows access from the old location.

This code largely just copies pointers around.

frankmcsherry3• 12 years ago

I totally left the #[inline(always)] out from the code, in the interest of explaining. That's what prevents the copying! :D

And also what Steve said about Rust being rly rly smart, or at least move semantics opening the door for that.

killercup• 12 years ago

There is a part 2: http://www.frankmcsherry.org/columnarization/serialization/r...

mkaufmann3• 12 years ago

While I like the construction of the column store and the corresponding API. The claims of the author don't really make sense:

"... columnarization, a technique from the database community for laying out structured records in a format that is more convenient for serialization than the records themselves."

Column stores in comparison to row stores don't offer any serialization benefit per se. The main benefits are the following, I will be using a record (A,B,C,D,E) as example with all types u32 (4 bytes):

* If you only use some fields you have to load less data from memory/disk into the CPU cache and your working set is more probable to fit into cache. For example when filtering only the records where A=22 and B=45 you only have to actually load x(sizeof(A)+sizeof(B)) = x8 bytes instead of xrecord_size=x20. This can make a very significant difference. * When using compression to reduce the size of data, columns can often be compressed better because they only contain data of the same type and nature and thus probably share similarities. When using such a small record consisting only of integers it probably won't make a difference. But if e.g. some fields are country abbreviations, textual description or others are ids, one could easily imagine that there are gains.

Coming back to the point about serialization, using the same technique as described in the blog post, there won't[1] be a performance difference between column storage and row storage (e.g. using a struct). The method described in the blog post just lets the data array of the original vector be wrapped by a Vec<u8> without even moving the memory, so the method is independent of the data type that is stored in the vectors. Of course it will only work for data types that do not contains references, otherwise we could get illegal memory access after deserialization (which should be guaranteed by the rust type system because only Copy types are allowed).

The only thing this benchmark is testing is how fast a vector can be initialized.

[1] There can be an space improvement of keeping the data in a column layout compared to row layout when using normal structs. Normal structs normally align the total size to the size of the largest field in the struct. A struct containing i64 and i8 would contain 7 bytes of padding. In a column layout this overhead would be avoided. Still there would not be an improvement in this serialization scheme as it does not actually copy any data.

frankmcsherry3• 12 years ago

Hi, author here.

Unless I misunderstand your post, I think you may have missed important parts!

More is going on than "just wrapping the data of the original vector as a Vec<u8>". That is what happens when we are handed a Vec<uint>, but for other inner types (pairs, vectors, etc) there is more to do (and, actual code presented showing what that work is).

1. A Vec<(u8, u64)> definitely ends up as a (Vec<u8>, Vec<u64>), thereby avoiding padding you'd have if you just wrote out the elements as structs. You absolutely end up saving space, and part of doing this is definitely copying data. (responding to "as it does not actually copy any data").

2. The types themselves can be vectors, corresponding to a struct with owned pointers (whose elements can also own pointers, etc). This is not something that just casting the source array will deal with. It's important here that Rust hands you ownership, as otherwise it would be totally inappropriate for us to claim the underlying memory (which the code does). That part, recycling owned memory, is one of the big performance wins (about 2.5x faster for me than invoking the allocator each time I need to mint a new array).

3. "Copy" doesn't appear in the first post. The types don't have to be Copy, and indeed Vec<T> is not Copy, even if T is. Not sure where you got that from. But, no such requirement; yay!

Read part 2 for more about Copy, and how when your type is Copy you get a free implementation that does what you suggest (keeping each struct intact), except you can mix and match with Vecs and Options and stuff like that.

Hope this clears up some of the not-sense-making. Feel free to holler with other questions.

Cheers, Frank

mkaufmann3• 12 years ago

Addressing your points:

1. In my comment I acknowledge that there are space savings with types that include padding. I don't understand why you imply in your answer that I didn't understand this point. Regarding my comment about not copying data: it was based on the benchmark[1] as linked in the blog post. The parts you measure do not copy any of the user data! It only converts your internal vectors of user types to a list of vectors of u8 type. So what you are doing is essentially moving the data. But when using move semantics the size of the user data does not matter any more, so there won't be a difference between your column layout and using the more complex types directly inside the vectors if measured the same way as in this benchmark. In my opinion your benchmark is flawed and does not support the argument you make in your blog post.

2. Ok yes that was bad wording on my side. If you provide adaptors for complex types with pointers like Vec than you can of course also serialize those.

3. I guess this is just about the same argument as point 2) and thus redundant.

I made the effort for you to write a serialization framework which does not do "columnarization" but which simulates row layout and also added an adaptor for vec and ran it with the same benchmark, and also reran the original columnar benchmark on my machine. Both benchmarks were compiled exactly the same way. You can find my code here[2]. Here are the results:

==============================

columnarization <uint> 12.1 GB/s, 742k values/s

columnarization <(uint, (uint, uint))> 5.3 GB/s, 107k values/s

columnarization <vec<uint>> 2.1 GB/s, 54k values/s

columnarization <Option<uint>> 1.58 GB/s, 163k values/s

==============================

row (simple_serialize) <uint> 12.1 GB/s, 730k values/s

row (simple_serialize) <(uint, (uint, uint))> 8.83 GB/s, 164 values/s

row (simple_serialize) <vec<uint>> 2.13 GB/s, 56 values/s

row (simple_serialize) <Option<uint>> 8.82 GB/s, 263k values/s

==============================

You see that columnization does not have a performance benefit in your benchmark and it even is significantly slower for Option<uint> type and pairs.

In your blog post you never mentioned that columnization will only has the potential to bring performance benefits to de-/serialization when using types with large padding overheads. I think this discussion would probably helped the blog post. It would be much better if it would either omit the currently wrong performance argument and just focus on the nicely typed API or if it would use a proper benchmark which would support your argument, which from my understanding would be only possible in a very limited set of use case scenarios.

Here is a list of other valid arguments you could have made instead:

* Format saves space at the cost of performance.

* Better than repr(packed) as it will also work on platforms that don't support unaligned access

[1] https://github.com/frankmcsherry/columnar/blob/master/exampl... [2] https://github.com/mkaufmann/simple_serialize

ww520• 12 years ago

Pretty cool. Columnarization is used in other contexts like high performance computing, or Entity Component System in gaming. It's data centric and very cache friendly.

Columnarization in Rust

Related Stories