Apart from the Steindorfer and Vinju paper [0], does anybody have any resources for implementing the CHAMP variant of HAMT tries?

Better yet, has anybody here implemented one before, and might I pick your brain?

[0] https://michael.steindorfer.name/publications/oopsla15.pdf

I've implemented CHAMP [0] for Peergos in the IPLD/IPFS setting, largely based on the Steindorfer paper. There is one improvement that whyrusleeping from ipfs came up with which is to allow a small number of hash collisions in a level, before pushing things down a level.

[0] https://github.com/Peergos/Peergos/blob/master/src/peergos/s...

Thanks for the link, this is super helpful.

I posted my parent comment opportunistically, so I have yet to distill my reading into specific, well-formed questions. Is there any way I can message you when I get around to doing so?

In the meantime:

>There is one improvement that whyrusleeping from ipfs came up with which is to allow a small number of hash collisions in a level, before pushing things down a level.

How is this an improvement, exactly?

Sure, I have a protonmail email with my github username.

>How is this an improvement, exactly? This allows you to essentially make the tree "fatter" for a given bitwidth. This matters more in the ipfs setting because links are not memory pointers, but Merkle-links to objects which may be a network request away.

Ah ok, that makes sense.

I've sent you an email with from my personal gmail account. I really appreciate the help, thanks!

Haven't implemented it myself, though I mean to try one day. But I've spoken to Michael before and used his implementation: https://github.com/usethesource/capsule

Oh wow! I imagine he's quite busy -- how did you get in touch with him. I know this is a long-shot, but any chance you might be able to arrange an email intro?

Happenstance really, he briefly joined our research group. I'll send him a link to this thread.

I have done more than enough work on the CHAMP encoded hashmap in Scala 2.13 to call myself a co-author

As I mentioned in another comment, I posted my parent comment opportunistically, so I have yet to distill my reading into specific, well-formed questions.

Is there any way I can message you when I get around to doing so? I'd really appreciate a super-quick whiteboard session, if you have the time.

Sure, my email is joshlemer [at] gmail [dot] com

Thanks! I've sent you an email.

Scala experimented[0] with implementing those in its standard library. I'm not sure if it made it in to the final release or not.

EDIT It looks like it did make it in: https://github.com/scala/scala/commit/d5ae93e1b35fac4002cfbf...

[0] https://github.com/scala/collection-strawman/issues/192

There's a C++ implementation here:

https://github.com/arximboldi/immer

Immer is amazing! I would suggest also looking at his work with transducers in the atria c++ library. The talk he gave about them is what made transducers really click for me.

Does https://idea.popcount.org/2012-07-25-introduction-to-hamt/ count?

I was looking for more CHAMP-specific stuff, but I've bookmarked this nonetheless. It's always nice to look at something in a few different ways :)

I have implemented something like that in Pascal:

https://github.com/benibela/hamt

Clojure has an implementation of these in java.

Nitpick: Clojure does not use the new CHAMP encoding

Can someone explain the reasoning for using a HAMT here? I have played around with hashtables quite a bit, but altough HAMT's are cool and a lot more memory efficient, they are likely not faster and will at some point suffer from more severe memory fragmentation than hashtables.

Is the concern speed or memory? Because speed would kind of seem strange, since after changing to the new architecture, they should see a strong decrease in messages being exchanged anyways, right?

Or are go's maps inefficient? Genuinely curious

HAMT is excellent for a persistent dictionary. Not persistent as in write-to-disk, but as in immutable-yet-small.

It is not clear from the article that the author needed a persistent data structure.

It’s also, frankly, not clear they‘d benefit from using an HAMT at all: the only measurements they present are of their problem domain, not of any implementations—and many of the asymptotics they cite are also true of hashmaps.

If their initial hashmap approach did have O(N) inserts because they copied the whole map for “immutability” every time, then sure, HAMTs will beat the pants off of it—but did it?

Btw, can it be used to build a persistent dictionary that keeps the key in insertion-order?

Sure. Map the HAMT to an index into a persistent vector or something e.g. Scala's VectorMap and TreeSeqMap. You could also reuse the old "linked list map" idea, but go through the keys every time e.g. whenever you add a value to the map, add a triple of `(Some(last_key), value, None)` and update the last keys value such that the third item becomes the key of the new item.

>Sure. Map the HAMT to an index into a persistent vector or something e.g. Scala's VectorMap and TreeSeqMap

I do not have a persistent vector, only a HAMT.

Perhaps the HAMT becomes a standard trie if one sets hashcode(i) = i, and then it can be enumerated in ascending key order?

But the double-lookup in two data structures is unnecessary slow. The first HAMT could store (Index, Value) pairs and only use the "vector" for enumeration.

But I was hoping for a single data structure that can do lookups and ordered enumerations.

>. You could also reuse the old "linked list map" idea, but go through the keys every time e.g. whenever you add a value to the map, add a triple of `(Some(last_key), value, None)` and update the last keys value such that the third item becomes the key of the new item.

I have been using a linked list as sole storage so far. The lookup was O(n), which was fast enough with my data, but I wanted to improve it to be more scalable.

But enumerating the list is in reverse, is it not? Going from the newest key to the oldest key. I want it to go from the oldest to the newest key.

In any case it should be better to store pointers to HAMT nodes rather than keys.

> I do not have a persistent vector, only a HAMT.

Why not get a persistent vector?

> But the double-lookup in two data structures is unnecessary slow.

Not necessarily. You'd have 2xlog(n) and because you wouldn't be storing the data in the HAMT it could be denser / wider and thus better fit in cache. At least that's what happened with Raymond Hettinger's (non-persistent) naturally ordered dict: rather than a sparse array of buckets, the hashmap is a sparse array of indices and a dense array of buckets leading to much better memory density (even more so with an adaptive sparse array where the stored indices can be u8, u16 or usize depending on the number of items: most maps will only need a sparse array of u8 leading to even better memory behaviour for the most problematic part).

> But enumerating the list is in reverse, is it not? Going from the newest key to the oldest key. I want it to go from the oldest to the newest key.

Not if you have a doubly linked list, but that increases your costs in storage and update complexity

> In any case it should be better to store pointers to HAMT nodes rather than keys.

If you're implementing the entire datastructure then yes, I was rather assuming you were on the outside with an existing HAMT.

>Why not get a persistent vector?

I have not read about them. Might take a while to implement.

I use Pascal. But it is hard to find Pascal libraries for anything. Either they do not exist at all or have not been maintained for over a decade

So, if I want to use a data structure, I usually need to implement it myself

>Not necessarily. You'd have 2xlog(n) and because you wouldn't be storing the data in the HAMT it could be denser / wider and thus better fit in cache.

If I reuse the HAMT as persistent vector, it is not so good. There is no grouping, each node is allocated separately and could be anywhere in memory. One lookup could be five cache misses.

>At least that's what happened with Raymond Hettinger's (non-persistent) naturally ordered dict: rather than a sparse array of buckets, the hashmap is a sparse array of indices and a dense array of buckets leading to much better memory density (even more so with an adaptive sparse array where the stored indices can be u8, u16 or usize depending on the number of items: most maps will only need a sparse array of u8 leading to even better memory behaviour for the most problematic part).

I use such a hashmap, too, for the insertion order. (there are Pascal libraries with hashmaps, but even then I needed some weeks to customize it). But all indices there are 32-bit. Perhaps I will add the adaptiveness to it. 16-bit at least. 8-bit seems hardly worth it, at most it can save 256 bytes

>> But enumerating the list is in reverse, is it not? Going from the newest key to the oldest key. I want it to go from the oldest to the newest key.

>Not if you have a doubly linked list, but that increases your costs in storage and update complexity

But a doubly linked list would not be persistent.

Arbitrary many new nodes could point to one old node, but the old node can only point to one of them

HAMT are efficient for persistent data structures, compared to more usual trees.

They’re slower and less efficient than a mutable hashmap, but way more so if you need to keep the “old revisions” around.

> they are likely not faster

If they are cache friendly then they can be very fast. The papers by Phil Bagwell (may he rest in peace) are very rich. Lots of implementation detail. Check them out!

Maybe I'm just missing it, but I don't see in this description any explanation of how to handle hash collisions - ie, when two keys map to the same hash value. Do you iterate through an infinite series of secondary hash functions when you run out of distinguishing bits? Or do you bottom out at a table that you linearly scan?

Overall, it seems like the algorithm is "maintain a trie on the hash values, using an auxiliary bitmap and popcnt to determine the index within any trie node." Does that miss anything?