In this article, we show that key-value stores backed by a log-structured merge-tree (LSM-tree) exhibit an intrinsic tradeoff between lookup cost, update cost, and main memory footprint, yet all existing designs expose a suboptimal and difficult to tune tradeoff among these metrics. We pinpoint the problem to the fact that modern key-value stores suboptimally co-tune the merge policy, the buffer size, and the Bloom filters’ false-positive rates across the LSM-tree’s different levels. We present Monkey, an LSM-tree based key-value store that strikes the optimal balance between the costs of updates and lookups with any given main memory budget. The core insight is that worst-case lookup cost is proportional to the sum of the false-positive rates of the Bloom filters across all levels of the LSM-tree. Contrary to state-of-the-art key-value stores that assign a fixed number of bits-per-element to all Bloom filters, Monkey allocates memory to filters across different levels so as to minimize the sum of their false-positive rates. We show analytically that Monkey reduces the asymptotic complexity of the worst-case lookup I/O cost, and we verify empirically using an implementation on top of RocksDB that Monkey reduces lookup latency by an increasing margin as the data volume grows (50--80% for the data sizes we experimented with). Furthermore, we map the design space onto a closed-form model that enables adapting the merging frequency and memory allocation to strike the best tradeoff among lookup cost, update cost and main memory, depending on the workload (proportion of lookups and updates), the dataset (number and size of entries), and the underlying hardware (main memory available, disk vs. flash). We show how to use this model to answer what-if design questions about how changes in environmental parameters impact performance and how to adapt the design of the key-value store for optimal performance.

Today's data deluge enables organizations to collect massive data, and analyze it with an ever-increasing number of concurrent queries. Traditional data warehouses (DW) face a challenging problem in executing this task, due to their query-centric model: each query is optimized and executed independently. This model results in high contention for resources. Thus, modern DW depart from the querycentric model to execution models involving sharing of common data and work. Our goal is to show when and how a DW should employ sharing. We evaluate experimentally two sharing methodologies, based on their original prototype systems, that exploit work sharing opportunities among concurrent queries at run-time: Simultaneous Pipelining (SP), which shares intermediate results of common sub-plans, and Global Query Plans (GQP), which build and evaluate a single query plan with shared operators.First, after a short review of sharing methodologies, we show that SP and GQP are orthogonal techniques. SP can be applied to shared operators of a GQP, reducing response times by 20%-48% in workloads with numerous common sub-plans. Second, we corroborate previous results on the negative impact of SP on performance for cases of low concurrency. We attribute this behavior to a bottleneck caused by the push-based communication model of SP. We show that pull-based communication for SP eliminates the overhead of sharing altogether for low concurrency, and scales better on multi-core machines than push-based SP, further reducing response times by 82%-86% for high concurrency. Third, we perform an experimental analysis of SP, GQP and their combination, and show when each one is beneficial. We identify a trade-off between low and high concurrency. In the former case, traditional query-centric operators with SP perform better, while in the latter case, GQP with shared operators enhanced by SP give the best results.

In the last few years NAND flash storage has become more and more popular as price per GB and capacity both improve at exponential rates. Flash memory offers significant benefits compared to magnetic hard disk drives (HDDs) and DBMSs are highly likely to use flash as a general storage backend, either alone or in heterogeneous storage solutions with HDDs. Flash devices, however, respond quite differently than HDDs for common access patterns, and recent research shows a strong asymmetry between read and write performance. Moreover, flash storage devices behave unpredictably, showing a high dependence on previous IO history and usage patterns.In this paper we investigate how a DBMS can overcome these issues to take full advantage of flash memory as persistent storage. We propose new a flash aware data layout -append and pack -which stabilizes device performance by eliminating random writes. We assess the impact of append and pack on OLTP workload performance using both an analytical model and micro-benchmarks, and our results suggest that significant improvements can be achieved for real workloads.

The shift to multi-core and multi-socket hardware brings new challenges to database systems, as the software parallelism determines performance. Even though database systems traditionally accommodate simultaneous requests, a multitude of synchronization barriers serialize execution. Write-ahead logging is a fundamental, omnipresent component in ARIES-style concurrency and recovery, and one of the most important yet-to-be addressed potential bottlenecks, especially in OLTP workloads making frequent small changes to data. In this paper, we identify four loggingrelated impediments to database system scalability. Each issue challenges different level in the software architecture: (a) the high volume of small-sized I/O requests may saturate the disk, (b) transactions hold locks while waiting for the log flush, (c) extensive context switching overwhelms the OS scheduler with threads executing log I/Os, and (d) contention appears as transactions serialize accesses to in-memory log data structures. We demonstrate these problems and address them with techniques that, when combined, comprise a holistic, scalable approach to logging. Our solution achieves a 20-69% speedup over a modern database system when running log-intensive workloads, such as the TPC-B and TATP benchmarks, in a single-socket multiprocessor server. Moreover, it achieves log insert throughput over 2.2 GB/s for small log records on the single-socket server, roughly 20 times higher than the traditional way of accessing the log using a single mutex. Furthermore, we investigate techniques on scaling the performance of logging to multi-socket servers. We present a set of optimizations which partly ameliorate the latency penalty that comes with multi-socket hardware, and then we investigate the feasibility of applying a distributed log buffer design at the socket level.

Database engines face growing scalability challenges as core counts exponentially increase each processor generation, and the efficiency of synchronization primitives used to protect internal data structures is a crucial factor in overall database performance. The trade-offs between different implementation approaches for these primitives shift significantly with increasing degrees of available hardware parallelism. Blocking synchronization, which has long been the favored approach in database systems, becomes increasingly unattractive as growing core counts expose its bottlenecks. Spinning implementations improve peak system throughput by a factor of 2x or more for 64 hardware contexts, but suffer from poor performance under load.In this paper we analyze the shifting trade-off between spinning and blocking synchronization, and observe that the trade-off can be simplified by isolating the load control aspects of contention management and treating the two problems separately: spinning-based contention management and blocking-based load control. We then present a proof of concept implementation that, for high concurrency, matches or exceeds the performance of both user-level spinlocks and the pthread mutex under a wide range of load factors. INTRODUCTIONRecent shifts in computer architecture have resulted in systems containing multiple cores per chip, with core counts projected to double every two years for the foreseeable future. While multicore architectures make available an unprecedented degree of hardware parallelism, they also pose new challenges for database engine design. Increasing the number of concurrent threads puts pressure on internal database engine components and exposes new bottlenecks in the system [7]. Database systems have long depended on blocking synchronization primitives (supplied by the operating system) to manage contention because they are both effective and offer predictable performance over a wide range of system load factors. However, we find that the trade-offs between spinning and blocking primitives shift significantly with increasing degrees of available hardware parallelism. In particular, blocking primitives become unattractive because they result in low system utilization for the high core counts which now appear in commodity servers. By utilizing fully, the machine spinlocks improve performance by 2x or more, but suffer from unstable performance under load.The strengths and weaknesses of spinning and blocking as contention management strategies are well known, and several approaches have been proposed to address the weaknesses or balance their trade-offs. For example, mutex locks in certain operating systems make use of adaptive spinning to avoid the cost of context switches when the wait for a lock 1 is short, leading to performance competitive with spinlocks while avoiding most of the weaknesses of spinning. Other work suggests heuristics for optimizing the duration of spinning based on machine size and workload [3]. Recent research [6] has also addressed partially t...

No abstract

Data-intensive applications fueled the evolution of log structured merge (LSM) based key-value engines that employ the out-of-place paradigm to support high ingestion rates with low read/write interference. These benefits, however, come at the cost of treating deletes as a second-class citizen. A delete inserts a tombstone that invalidates older instances of the deleted key. State-of-the-art LSM engines do not provide guarantees as to how fast a tombstone will propagate to persist the deletion. Further, LSM engines only support deletion on the sort key. To delete on another attribute (e.g., timestamp), the entire tree is read and re-written. We highlight that fast persistent deletion without affecting read performance is key to support: (i) streaming systems operating on a window of data, (ii) privacy with latency guarantees on the right-to-be-forgotten, and (iii) en masse cloud deployment of data systems that makes storage a precious resource.To address these challenges, in this paper, we build a new key-value storage engine, Lethe, that uses a very small amount of additional metadata, a set of new delete-aware compaction policies, and a new physical data layout that weaves the sort and the delete key order. We show that Lethe supports any user-defined threshold for the delete persistence latency offering higher read throughput (1.17 − 1.4×) and lower space amplification (2.1 − 9.8×), with a modest increase in write amplification (between 4% and 25%). In addition, Lethe supports efficient range deletes on a secondary delete key by dropping entire data pages without sacrificing read performance nor employing a costly full tree merge.

No abstract