Comment by koverstreet
14 hours ago
No multithreaded write benchmarks. That's a major omission, given that's where you'll see the biggest difference between b-trees and LSM trees.
The paper also talks about the overhead of the mapping table for node lookups, and says "Bf-Tree by default pins the inner nodes in memory and uses direct pointer addresses to reference them. This allows a simpler inner node implementation, efficient node access, and reduced con- tention on the mapping table".
But you don't have to pin nodes in memory to use direct pointer lookups. Reserve a field in your key/value pair for a direct in-memory pointer, and after chasing it check that you got the node you expect; only fall back to the mapping table (i.e. hash table of cached nodes) if the pointer is uninitialized or you don't get the node you expect.
"For write, conventional B-Tree performs the worst, as a single record update would incur a full page write, as evidenced by the highest disk IO per operation."
Only with a random distribution, but early on the paper talks about benchmarking with a Zip-f distribution. Err?
The benchmark does look like a purely random distribution, which is not terribly realistic for most use cases. The line about "a single record updating incurring a full page write" also ignores the effect of cache size vs. working set size, which is a rather important detail. I can't say I trust the benchmark numbers.
Prefix compression - nice to see this popping up.
"hybrid latching" - basically, they're doing what the Linux kernel calls seqlocks for interior nodes. This is smart, but given that their b-tree implementation doesn't use it, you shouldn't trust the b-tree benchmarks.
However, I found that approach problematic - it's basically software transactional memory, with all the complexity that implies, and it bleeds out into too much of the rest of your b-tree code. Using a different type of lock for interior nodes where read locks only use percpu counters gives the same performance (read locks touch no cachelines shared by other CPUs) for much lower complexity.
Not entirely state of the art, and I see a lot of focus on optimizations that likely wouldn't survive in a larger system, but it does look like a real improvement over LSM trees.
Sure, but on principle, looking at the paper, I'd expect it to outperform B-trees since write amplification is reduced, generally. You thinking about cases requiring ordering of writes to a given record (lock contention)?
I think their claims of write amplification reduction are a bit overstated given more realistic workloads.
It is true that b-trees aren't ideal in that respect, and you will see some amount of write amplification, but not enough that it should be a major consideration, in my experience
You really have to take into account workingset size and cache size to make any judgements there; your b-tree writes should be given by journal/WAL reclaim, which will buffer up updates.
A purely random update workload will kill a conventional b-tree on write amplification - like I mentioned, that's the absolute worst case scenario for a b-tree. But it just doesn't happen in the real world.
For the data I can give you, that would be bcachefs's hybrid b-tree - large btree nodes (256k, typically) which are internally log structured; I would consider it a minor variation on a classical b-tree. The log structuring mean that we can incrementally write only the dirty keys in a node, at the cost of some compaction overhead (drastically less than a conventional LSM).
In actual real world usage, when I've looked at the numbers (not recently, so this may have changed) we're always able to do giant highly efficient b-tree writes - the journal and in-memory cache are batching things up as much as we want - which means write amplification is negligible.
Also you can use dense B+-Trees for reads possibly with some bloom filters or the like if you expect/profile a high fraction of negative lookups, use LSM to eventually compact, and get both SSD/ZNS friendly write patterns as well as full freedom to only compact a layer once it's finer state is no longer relevant to any MVCC/multi-phase-commit schemes. Being able to e.g. run a compression algorithm until you just exceed the storage page size, take it's state from just before it exceeded, and begin the next bundle with the entry that made you exceed the page size.... It's quite helpful when storage space or IO bandwidth is somewhat scarce.
If you're worried about the last layer being a giant unmanageably large B+-Tree, just shard it similarly in key space to not need much free temporary working space on SSD to stream the freshly compacted data to while the inputs to the compaction still serve real time queries.
Of course mileage may vary with different workloads, but are there any good benchmarks/suites to use for comparison in cases like these? They used YCSB but I don't know if those workloads ([1]) are relevant to modern/typical access patterns nor if they're applicable to SQL databases.
You thinking about running some benchmarks in a bcachefs branch (:pray:)?
I want to see this data structure prototyped in PostgreSQL.
[1]: https://github.com/brianfrankcooper/YCSB/tree/master/workloa...
2 replies →