Biblio

Non-volatile memory express (NVMe) based SSDs and the NUMA platform are widely adopted in servers to achieve faster storage speed and more powerful processing capability. As of now, very little research has been conducted to investigate the performance and energy efficiency of the stateof-the-art NUMA architecture integrated with NVMe SSDs, an emerging technology used to host parallel I/O threads. As this technology continues to be widely developed and adopted, we need to understand the runtime behaviors of such systems in order to design software runtime systems that deliver optimal performance while consuming only the necessary amount of energy. This paper characterizes the runtime behaviors of a Linuxbased NUMA system employing multiple NVMe SSDs. Our comprehensive performance and energy-efficiency study using massive numbers of parallel I/O threads shows that the penalty due to CPU contention is much smaller than that due to remote access of NVMe SSDs. Based on this insight, we develop a dynamic “lesser evil” algorithm called ESN, to minimize the impact of these two types of penalties. ESN is an energyefficient profiling-based I/O thread scheduler for managing I/O threads accessing NVMe SSDs on NUMA systems. Our empirical evaluation shows that ESN can achieve optimal I/O throughput and latency while consuming up to 50% less energy and using fewer CPUs.

Recent studies have found that parallel garbage collection performs worse with more CPUs and more collector threads. As part of this work, we further investigate this enomenon and find that poor scalability is worst in highly scalable Java applications. Our investigation to find the causes clearly reveals that efficient multi-threading in an application can prolong the average object lifespan, which results in less effective garbage collection. We also find that prolonging lifespan is the direct result of Linux's Completely Fair Scheduler due to its round-robin like behavior that can increase the heap contention between the application threads. Instead, if we use pseudo first-in-first-out to schedule application threads in large multicore systems, the garbage collection scalability is significantly improved while the time spent in garbage collection is reduced by as much as 21%. The average execution time of the 24 Java applications used in our study is also reduced by 11%. Based on this observation, we propose two approaches to optimally select scheduling policies based on application scalability profile. Our first approach uses the profile information from one execution to tune the subsequent executions. Our second approach dynamically collects profile information and performs policy selection during execution.

Parallel garbage collection has been used to speedup the collection process on multicore architectures. Similar to other parallel techniques, balancing the workload among threads is critical to ensuring good overall collection performance. To this end, work stealing is employed by the current stateof-the-art Java Virtual Machine, OpenJDK, to keep GC threads from idling during a collection process. However, we found that the current algorithm is not efficient. Its usage can often cause GC performance to be worse than when work stealing is not used. In this paper, we identify three factors that affect work stealing efficiency: determining tasks that can benefit from stealing, frequency with which to attempt stealing, and performance impacts of failed stealing attempts. Based on this analysis, we propose SmartStealing, a new algorithm that can automatically decide whether to attempt stealing at a particular point during execution. If stealing is attempted, it can efficiently identify a task to steal from. We then compare the collection performances when (i) the default work stealing algorithm is used, (ii) work stealing is not used at all, and (iii) the SmartStealing approach is used. Without modifying the remaining garbage collection system, the evaluation result shows that SmartStealing can reduce the parallel GC execution time for 19 of the 21 benchmarks. The average reduction is 50.4% and the highest reduction is 78.7%. We also investigate the performances of SmartStealing on NUMA and UMA architectures.