Information grows monotonically.
Kudzu provides a Map
and Set
implemented on top of a concurrent skiplist.
The key difference between the types in kudzu and other concurrent data
structures is that kudzu's data structures do not support remove operations.
This limitation makes the types much simpler to implement and hopefully more
performant, with less coordination overhead, while still being useful for many
applications.
These can be used for any concurrent algorithm in which a map or set only grows, without ever losing members. For example, this can be combined with rayon as a memoization table for divide-and-conquer algorithms with repeating subproblems (e.g. fibonacci).
Assuming my implementation is correct and my reasoning is sound (big assumptions), all of these things should be true:
- Data races are impossible.
- All operations on the map are lock-free: it is not possible for two concurrent operations on the same set to create a deadlock.
- Lookup is wait-free: looking up an item in the set never waits on another thread to complete.
- At no point is the skiplist in a state inconsistent with the properties of a skiplist (that is, every lane will always be a subset of the lane below it).
Lookup accesses are performed by simply searching through the skiplist using Relaxed loads, with no locking or coordination.
Inserts are performed by finding the location to insert and performing a CAS on the pointers in each lane that the node will be inserted to. If the CAS fails on the lowest lane, the node is not inserted and the entire insertion algorithm is tried again. However, on any higher lane than the lowest, we simply stop inserting once an insertion fails. This leaves the list flatter than it would ideally be in instances of contention, but makes handling that contention much cheaper than retrying would be - I hope this proves to be a beneficial trade off.
(Inserting an element which is found to already be present returns the element you attempted to insert without changing the value already in the set.)
Because of this insertion strategy and the absence of removal operations, concurrency correctness can be maintained by simply using atomic CAS operations on each modified pointer, rather than having to track additional metadata and node-level locks.
I owe much to this paper, which describes a somewhat similar algorithm and suggests something similar to my algorithm in the conclusion.
This skiplist also has a highly optimized memory layout to improve performance. Each node contains all of its lanes inline with the element, but is also precisely sized so that it does not contain space for unused lanes. This means that the space overhead for each node is only on average 2 pointers and 1 byte (to track the number of lanes for deallocation purposes).
We also store lanes in reverse order, with the highest lane as the first element. Between this and making the nodes inline, we should have much better memory locality, because nodes are most often visited in their highest lane, not their lowest.