Translation Look aside Buffers (TLBs) are ubiquitously used in modern architectures to cache virtual-to-physical mappings and, as they are looked up on every memory access, are paramount to performance scalability. The emergence of chip-multiprocessors (CMPs) with per-core TLBs, has brought the problem of TLB coherence to front stage. TLBs are kept coherent at the software-level by the operating system (OS). Whenever the OS modifies page permissions in a page table, it must initiate a coherency transaction among TLBs, a process known as a TLB shoot down. Current CMPs rely on the OS to approximate the set of TLBs caching a mapping and synchronize TLBs using costly Inter-Proceessor Interrupts (IPIs) and software handlers. In this paper, we characterize the impact of TLB shoot downs on multiprocessor performance and scalability, and present the design of a scalable TLB coherency mechanism. First, we show that both TLB shoot down cost and frequency increase with the number of processors and project that software-based TLB shoot downs would thwart the performance of large multiprocessors. We then present a scalable architectural mechanism that couples a shared TLB directory with load/store queue support for lightweight TLB invalidation, and thereby eliminates the need for costly IPIs. Finally, we show that the proposed mechanism reduces the fraction of machine cycles wasted on TLB shoot downs by an order of magnitude.
Library OSs have been proposed to deploy applications isolated inside containers, VMs, or trusted execution environments. They often follow a highly modular design in which third-party components are combined to offer the OS functionality needed by an application, and they are customised at compilation and deployment time to fit application requirements. Yet their monolithic design lacks isolation across components: when applications and OS components contain security-sensitive data (e.g., cryptographic keys or user data), the lack of isolation renders library OSs open to security breaches via malicious or vulnerable third-party components. We describe CubicleOS, a library OS that isolates components in the system while maintaining the simple, monolithic development approach of library composition. CubicleOS allows isolated components, called cubicles , to share data dynamically with other components. It provides spatial memory isolation at the granularity of function calls by using Intel MPK at user-level to isolate components. At the same time, it supports zero-copy data access across cubicles with feature-rich OS functionality. Our evaluation shows that CubicleOS introduces moderate end-to-end performance overheads in complex applications: 2× for the I/O-intensive NGINX web server with 8 partitions, and 1.7–8× for the SQLite database engine with 7 partitions.
The increasing number of cores in manycore architectures causes important power and scalability problems in the memory subsystem. One solution is to introduce scratchpad memories alongside the cache hierarchy, forming a hybrid memory system. Scratchpad memories are more power-efficient than caches and they do not generate coherence traffic, but they suffer from poor programmability. A good way to hide the programmability difficulties to the programmer is to give the compiler the responsibility of generating code to manage the scratchpad memories. Unfortunately, compilers do not succeed in generating this code in the presence of random memory accesses with unknown aliasing hazards.\ud This paper proposes a coherence protocol for the hybrid memory system that allows the compiler to always generate code to manage the scratchpad memories. In coordination with the compiler, memory accesses that may access stale copies of data are identified and diverted to the valid copy of the data. The proposal allows the architecture to be exposed to the programmer as a shared memory manycore, maintaining the programming simplicity of shared memory models and preserving backwards compatibility. In a 64-core manycore, the coherence protocol adds overheads of 4% in performance, 8% in network traffic and 9% in energy consumption to enable the usage of the hybrid memory system that, compared to a cache-based system, achieves a speedup of 1.14x and reduces on-chip network traffic and energy consumption by 29% and 17%, respectively.This work has been supported by the Spanish Government (grant SEV-2011-00067 of the Severo Ochoa Program), by the Spanish Ministry of Science and Innovation (contract\ud TIN2012-34557), by Generalitat de Catalunya (contracts 2014-SGR-1051 and 2014-SGR-1272), and by the RoMoL ERC Advanced Grant (GA 321253). Miquel Moreto has been partially supported by the Ministry of Economy and Competitiveness under Juan de la Cierva postdoctoral fellowship number JCI-2012-15047, and Marc Casas is supported by the Secretary for Universities and Research of the Ministry of Economy and Knowledge of the Government of Catalonia and the Cofund programme of the Marie Curie Actions of the 7th R&D Framework Programme of the European Union (Contract 2013 BP_B 00243).Postprint (author's final draft
As a basic building block of many applications, sorting algorithms that efficiently run on modern machines are key for the performance of these applications. With the recent shift to using GPUs for general purpose compuing, researches have proposed several sorting algorithms for single-GPU systems. However, some workstations and HPC systems have multiple GPUs, and applications running on them are designed to use all available GPUs in the system.\ud In this paper we present a high performance multi-GPU merge sort algorithm that solves the problem of sorting data distributed across several GPUs. Our merge sort algorithm first sorts the data on each GPU using an existing single-GPU sorting algorithm. Then, a series of merge steps produce a globally sorted array distributed across all the GPUs in the system. This merge phase is enabled by a novel pivot selection algorithm that ensures that merge steps always distribute data evenly among all GPUs. We also present the implementation of our sorting algorithm in CUDA, as well as a novel inter-GPU communication technique that enables this pivot selection algorithm. Experimental results show that an efficient implementation of our algorithm achieves a speed up of 1.9x when running on two GPUs and 3.3x when running on four GPUs, compared to sorting on a single GPU. At the same time, it is able to sort two and four times more records, compared to sorting on one GPU.Peer ReviewedPostprint (published version