Relativistic programming

Abstract: This paper presents a novel concurrent hash table implementation which supports wait-free, near-linearly scalable lookup, even in the presence of concurrent modifications. In particular, this hash table implementation supports concurrent moves of hash table elements between buckets, for purposes such as renames.Implementation of this algorithm in the Linux kernel demonstrates its performance and scalability. Benchmarks on a 64-way POWER system showed a 6x scalability improvement versus fine-grained locking, and a 1.5x improvement versus the current state of the art in Linux.To achieve these scalability improvements, the hash table implementation uses a new concurrent programming technique known as relativistic programming. This approach uses a copy-based update strategy to allow readers and writers to run concurrently without conflicts, avoiding many of the non-scalable costs of synchronization, inter-processor communication, and cache coherence. New techniques such as the proposed hash-table move algorithm allow readers to tolerate the resulting weak memory-ordering behavior that arises from allowing one version of a structure to be read concurrently with updates to a different version of the same structure. Relativistic programming techniques provide performance and scalability advantages over traditional synchronization, as demonstrated through several benchmarks.

Abstract: This paper presents algorithms for concurrently reading and modifying a red-black tree RBTree. The algorithms allow wait-free, linearly scalable lookups in the presence of concurrent inserts and deletes. They have deterministic response times for a given tree size and uncontended read performance that is at least 60% faster than other known approaches. The techniques used to derive these algorithms arise from a concurrent programming methodology called relativistic programming. Relativistic programming introduces write-side delay primitives that allow the writer to pay most of the cost of synchronization between readers and writers. Only minimal synchronization overhead is placed on readers. Relativistic programming avoids unnecessarily strict ordering of read and write operations while still providing the capability to enforce linearizability. This paper shows how relativistic programming can be used to build a concurrent RBTree with synchronization-free readers and both lock-based and transactional memory-based writers. Copyright © 2013 John Wiley & Sons, Ltd.

Abstract: Relativistic Programming is a technique that allows low overhead, linearly-scalable concurrent reads. It also allows joint access parallelism between readers and a writer. Unfortunately, it has so far been limited to a single writer so it does not scale on the write side. Software Transactional Memory (STM) is a technique that allows programs to take advantage of disjoint access parallelism on both the read-side and write-side. Unfortunately, STM systems have a higher overhead than many other synchronization mechanisms so although STM scales, STM starts from a lower baseline. We propose combining relativistic programming and software transactional memory in a way that gets the best of both worlds: low-overhead linearly-scalable reads that never conflict with writes and scalable disjoint access parallel writes. We argue for the correctness of our approach and present performance data that shows an actual implementation that delivers the promised performance characteristics.

Abstract: We present algorithms for shrinking and expanding a hash table while allowing concurrent, wait-free, linearly scalable lookups. These resize algorithms allow the hash table to maintain constant-time performance as the number of entries grows, and reclaim memory as the number of entries decreases, without delaying or disrupting readers. We implemented our algorithms in the Linux kernel, to test their performance and scalability. Benchmarks show lookup scalability improved 125x over readerwriter locking, and 56% over the current state-of-the-art for Linux, with no performance degradation for lookups during a resize. To achieve this performance, this hash table implementation uses a new concurrent programming methodology known as relativistic programming. In particular, we make use of an existing synchronization primitive which waits for all current readers to finish, with little to no reader overhead; careful use of this primitive allows ordering of updates without read-side synchronization or memory barriers.