memtable/skiplist: add a purpose-built skiplist

[ajwerner]

Fixes #95.

Supersedes #98

[marvin-j97]

With the Borrow<Q> trait bound it's not possible to use InternalKeyRef that would allow to not build an InternalKey (which involves a heap allocation).
Need to use equivalent::Comparable instead.

See: #98

[marvin-j97]

Need to play with this a bit - but should probably be much higher by default: 1 MB or so?

[ajwerner]

Yeah, it should be bigger than 32k, but 1MiB might be too big. The keys and values are not inline, it’s just the metadata. The questions I’d have are how expensive is allocating a new block, and how expensive is inserting into the skip map. My guess is that the alloc is not likely worse than 10us (it’s probably way less) and the inserts are ~100ns. If you can fit 1000 in here (if we say the average links is 32 and the key and value are each 32 bytes), then you’ll have spent at least 10x as long doing the inserting. In practice I think the mallocs even with zeroing is a lot cheaper. The benchmarks I was playing with don’t show much win above 256KiB.

[marvin-j97]

Eventually, for write transactions, the size should be much smaller (so that small transactions don't overallocate too much) - so this needs to be a non-generic parameter.

[ajwerner]

Okay, I can do that.

[ajwerner]

I’ll play with updating. It’s pretty hard to not leak the node from the previous update without hooking up a free list but it’s also not so hard to add one.

[ajwerner]

Okay, I can do that.

[ajwerner]

Yeah, it should be bigger than 32k, but 1MiB might be too big. The keys and values are not inline, it’s just the metadata. The questions I’d have are how expensive is allocating a new block, and how expensive is inserting into the skip map. My guess is that the alloc is not likely worse than 10us (it’s probably way less) and the inserts are ~100ns. If you can fit 1000 in here (if we say the average links is 32 and the key and value are each 32 bytes), then you’ll have spent at least 10x as long doing the inserting. In practice I think the mallocs even with zeroing is a lot cheaper. The benchmarks I was playing with don’t show much win above 256KiB.

[marvin-j97]

Todo:

• Evaluate memory overhead vs crossbeam skiplist
• Evaluate read and write latency vs crossbeam skiplist
• Try in a full system (heavily cached) benchmark (https://github.com/marvin-j97/rust-storage-bench) vs crossbeam skiplist

[marvin-j97]

I've tried rebasing this onto the v3 branch.

Outlook now: we don't need a free list because all insertions will be unique (because of unique sequence number).

[marvin-j97]

Removing the const generic N parameter is actually a bit harder than expected because it defines the layout of Buffer; changing it to static makes pointer derefs fail with misaligned pointer dereference.