Multicore Scalability of Concurrent Objects

We study the design, implementation, performance, and scalability of concurrent objects on multicore systems. In the first part we analyze the apparent trade-off between adherence to concurrent data structure semantics and scalability based on a concurrent FIFO queue as running example. We quantitatively relax the sequential specification of a FIFO queue to a k-FIFO queue, which can be understood as queue where each element may be dequeued out-of-order up to k − 1 or as pool where each element is dequeued within a k-bounded number of dequeue operations. We present different k-FIFO queue implementations with constant k, workload-dependent k, and unbounded k, and show experimentally that there exist optimal and robust k that result in best performance and scalability. Moreover, we introduce the notion of semantical deviation to quantify the difference in semantics between a run of a queue with relaxed FIFO semantics and a run of a regular FIFO queue on a given workload. We show in an extensive experimental evaluation that our k-FIFO queue implementations outperform and outscale many state-of-the-art concurrent queue and pool algorithms on different workloads and relate the performance results to the corresponding measured semantical deviations. In the second part we focus on concurrent heap management systems. Firstly, we present the design and implementation of a real-time memory allocator called Compact-fit (CF) which provides temporal and spatial guarantees. We analyze the throughput, latency, and scalability versus memory fragmentation trade-off of CF in experiments. Secondly, we introduce hierarchical allocation buffers (HABs), a memory partitioning scheme for highthroughput heap management systems. We present an implementation of HABs in the Hotspot Java virtual machine and analyze the trade-off between throughput and scalability on one hand, and memory fragmentation on the other hand.