Engineering Papers⌕ Search

SEARCH · Engineering Papers

Results for “cache locality”

Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 91 records · Page 5

Multiple core computer processor with globally-accessible local memories

A multi-core computer processor including a plurality of processor cores interconnected in a Network-on-Chip (NoC) architecture, a plurality of caches, each of the plurality of caches being associated with one and only one of the plurality of processor cores, and a plurality of memories, each of the plurality of memories being associated with a different set of at least one of the plurality of processor cores and each of the plurality of memories being configured to be visible in a global memory address space such that the plurality of memories are visible to two or more of the plurality of processor cores.

97 MATHEMATICS AND COMPUTING↗

Cloud services for the Fermilab scientific stakeholders

As part of the Fermilab/KISTI cooperative research project, Fermilab has successfully run an experimental simulation workflow at scale on a federation of Amazon Web Services (AWS), FermiCloud, and local FermiGrid resources. We used the CernVM-FS (CVMFS) file system to deliver the application software. We established Squid caching servers in AWS as well, using the Shoal system to let each individual virtual machine find the closest squid server. We also developed an automatic virtual machine conversion system so that we could transition virtual machines made on FermiCloud to Amazon Web Services. We used this system to successfully run a cosmic ray simulation of the NOvA detector at Fermilab, making use of both AWS spot pricing and network bandwidth discounts to minimize the cost. On FermiCloud we also were able to run the workflow at the scale of 1000 virtual machines, using a private network routable inside of Fermilab. As a result, we present in detail the technological improvements that were used to make this work a reality.

97 MATHEMATICS AND COMPUTING↗

CDF software distribution on the grid using Parrot

Large international collaborations that use decentralized computing models are becoming a custom rather than an exception in High Energy Physics. A good computing model for such big collaborations has to deal with the distribution of the experiment-specific software around the world. When the CDF experiment developed its software infrastructure, most computing was done on dedicated clusters. As a result, libraries, configuration files and large executables were deployed over a shared file system. In order to adapt its computing model to the Grid, CDF decided to distribute its software to all European Grid sites using Parrot, a user-level application capable of attaching existing programs to remote I/O systems through the filesystem interface. This choice allows CDF to use just one centralized source of code and a scalable set of caches all around Europe to efficiently distribute its code and requires almost no interaction with the existing Grid middleware or with local system administrators. This system has been in production at CDF in Europe since almost two years. Here, we present CDF implementation of Parrot and some comments on its performances.

99 GENERAL AND MISCELLANEOUS↗

Quantum Dynamics at Scale: Ultrafast Control of Emergent Functional Materials

Confluence of extreme-scale quantum dynamics simulations (i.e. quantum@scale) and cutting-edge x-ray free-electron laser experiments are revolutionizing materials science. An archetypal example is the exciting concept of using picosecond light pulses to control emergent material properties on demand in atomically-thin layered materials. This paper describes efforts to scale our quantum molecular dynamics engine toward the United States' first exaflop/s computer, under an Aurora Early Science Program project named "Metascalable layered material genome". Key algorithmic and computing techniques incorporated are: (1) globally-scalable and locally-fast solvers within a linear-scaling divide-conquer-recombine algorithmic framework; (2) algebraic `BLASification' of computational kernels; and (3) data alignment and loop restructuring, along with register and cache blocking, for enhanced vectorization and efficient memory access. The resulting weak-scaling parallel efficiency was 0.93 on 131,072 Intel Xeon Phi cores for a 56.6 million atom (or 169 million valence-electron) system, whereas the various code transformations achieved 5-fold speedup. The optimized simulation engine allowed us for the first time to establish a significant effect of substrate on the dynamics of layered material upon electronic excitation.

Algebraic BLASification↗

Locality Aware Concurrent Start for Stencil Applications

Stencil computations are at the heart of many physical simulations used in scientific codes. Thus, there exists a plethora of optimization efforts for this family of computations. Among these techniques, tiling techniques that allow concurrent start have proven to be very efficient in providing better performance for these critical kernels. Nevertheless, with many core designs being the norm, these optimization techniques might not be able to fully exploit locality (both spatial and temporal) on multiple levels of the memory hierarchy without compromising parallelism. It is no longer true that the machine can be seen as a homogeneous collection of nodes with caches, main memory and an interconnect network. New architectural designs exhibit complex grouping of nodes, cores, threads, caches and memory connected by an ever evolving network-on-chip design. These new designs may benefit greatly from carefully crafted schedules and groupings that encourage parallel actors (i.e. threads, cores or nodes) to be aware of the computational history of other actors in close proximity. In this paper, we provide an efficient tiling technique that allows hierarchical concurrent start for memory hierarchy aware tile groups. Each execution schedule and tile shape exploit the available parallelism, load balance and locality present in the given applications. We demonstrate our technique on the Intel Xeon Phi architecture with selected and representative stencil kernels. We show improvement ranging from 5.58% to 31.17% over existing state-of-the-art techniques.

Locality Aware execution, jagged tiling↗

Improved cache performance in Monte Carlo transport calculations using energy banding

We present an energy banding algorithm for Monte Carlo (MC) neutral particle transport simulations which depend on large cross section lookup tables. In MC codes, read-only cross section data tables are accessed frequently, exhibit poor locality, and are typically too much large to fit in fast memory. Thus, performance is often limited by long latencies to RAM, or by off-node communication latencies when the data footprint is very large and must be decomposed on a distributed memory machine. The proposed energy banding algorithm allows maximal temporal reuse of data in band sizes that can flexibly accommodate different architectural features. The energy banding algorithm is general and has a number of benefits compared to the traditional approach. In the present analysis we explore its potential to achieve improvements in time-to-solution on modern cache-based architectures.

97 MATHEMATICS AND COMPUTING↗

Exploring Hardware Support For Scaling Irregular Applications on Multi-node Multi-core Architectures

With the recent emergence of large-scale knowledge dis- covery, data mining and social network analysis, irregular applications have gained renewed interest. Classic cache-based high-performance architectures do not provide optimal performances with such kind of workloads, mainly due to the very low spatial and temporal locality of the irregular control and memory access patterns. In this paper, we present a multi-node, multi-core, fine-grained multi-threaded shared-memory system architecture specifically designed for the execution of large-scale irregular applications, and built on top of three pillars, that we believe are fundamental to support these workloads. First, we offer transparent hardware support for Partitioned Global Address Space (PGAS) to provide a large globally-shared address space with no software library overhead. Second, we employ multi-threaded multi-core processing nodes to achieve the necessary latency tolerance required by accessing global memory, which potentially resides in a remote node. Finally, we devise hardware support for inter-thread synchronization on the whole global address space. We first model the performances by using an analytical model that takes into account the main architecture and application characteristics. We describe the hardware design of the proposed cus- tom architectural building blocks that provide support for the above- mentioned three pillars. Finally, we present a limited-scale evaluation of the system on a multi-board FPGA prototype with typical irregular kernels and benchmarks. The experimental evaluation demonstrates the architecture performance scalability for different configurations of the whole system.

FPGA↗

Performance Impact and Trade-Offs for Tuning Key Architectural Parameters on CPU+GPU Systems

In this work, we performed an initial design space exploration of an accelerated processing unit (APU)—a hybrid CPU+GPU architecture that integrates both compute units (CUs) and memory into a unified system. This integration aims to reduce data movement, enhance memory locality, and improve energy efficiency by enabling the CPU and GPU to share memory directly. This effort focused on the interplay of key design components—cache line size, the number of CUs, and main memory technology—and the trade-offs of each configuration were analyzed. This paper highlights the various configurations’ impact on memory accesses, data reuse, and power utilization. The results provide valuable insights that can be leveraged to optimize APU architectures for high-performance and energy-efficient computing and thus create a balanced architecture. This optimization can be achieved by adopting dynamic cache management, runtime CU scaling, and advanced memory integration, highlighting the potential of APUs to address critical challenges in compute, data movement, and memory power consumption.

Asifuzzaman, Kazi [ORNL] (ORCID:0000000240044791)↗

Flexible Data-Aware Scheduling for Workflows over an In-Memory Object Store

This paper explores novel techniques for improving the performance of many-task workflows based on the Swift scripting language. We propose novel programmer options for automated distributed data placement and task scheduling. These options trigger a data placement mechanism used for distributing intermediate workflow data over the servers of Hercules, a distributed key-value store that can be used to cache file system data. We demonstrate that these new mechanisms can significantly improve the aggregated throughput of many-task workflows with up to 86x, reduce the contention on the shared file system, exploit the data locality, and trade off locality and load balance.

Duro, Francisco Rodrigo↗

Multi-core and many-core shared-memory parallel raycasting volume rendering optimization and tuning

Given the computing industry trend of increasing processing capacity by adding more cores to a chip, the focus of this work is tuning the performance of a staple visualization algorithm, raycasting volume rendering, for shared-memory parallelism on multi-core CPUs and many-core GPUs. Our approach is to vary tunable algorithmic settings, along with known algorithmic optimizations and two different memory layouts, and measure performance in terms of absolute runtime and L2 memory cache misses. Our results indicate there is a wide variation in runtime performance on all platforms, as much as 254% for the tunable parameters we test on multi-core CPUs and 265% on many-core GPUs, and the optimal configurations vary across platforms, often in a non-obvious way. For example, our results indicate the optimal configurations on the GPU occur at a crossover point between those that maintain good cache utilization and those that saturate computational throughput. This result is likely to be extremely difficult to predict with an empirical performance model for this particular algorithm because it has an unstructured memory access pattern that varies locally for individual rays and globally for the selected viewpoint. Our results also show that optimal parameters on modern architectures are markedly different from those in previous studies run on older architectures. In addition, given the dramatic performance variation across platforms for both optimal algorithm settings and performance results, there is a clear benefit for production visualization and analysis codes to adopt a strategy for performance optimization through auto-tuning. These benefits will likely become more pronounced in the future as the number of cores per chip and the cost of moving data through the memory hierarchy both increase.

97 MATHEMATICS AND COMPUTING↗

Multi-core and Many-core Shared-memory Parallel Raycasting Volume Rendering Optimization and Tuning

Given the computing industry trend of increasing processing capacity by adding more cores to a chip, the focus of this work is tuning the performance of a staple visualization algorithm, raycasting volume rendering, for shared-memory parallelism on multi-core CPUs and many-core GPUs. Our approach is to vary tunable algorithmic settings, along with known algorithmic optimizations and two different memory layouts, and measure performance in terms of absolute runtime and L2 memory cache misses. Our results indicate there is a wide variation in runtime performance on all platforms, as much as 254% for the tunable parameters we test on multi-core CPUs and 265% on many-core GPUs, and the optimal configurations vary across platforms, often in a non-obvious way. For example, our results indicate the optimal configurations on the GPU occur at a crossover point between those that maintain good cache utilization and those that saturate computational throughput. This result is likely to be extremely difficult to predict with an empirical performance model for this particular algorithm because it has an unstructured memory access pattern that varies locally for individual rays and globally for the selected viewpoint. Our results also show that optimal parameters on modern architectures are markedly different from those in previous studies run on older architectures. And, given the dramatic performance variation across platforms for both optimal algorithm settings and performance results, there is a clear benefit for production visualization and analysis codes to adopt a strategy for performance optimization through auto-tuning. These benefits will likely become more pronounced in the future as the number of cores per chip and the cost of moving data through the memory hierarchy both increase.

97 MATHEMATICS AND COMPUTING↗

Feedback-directed page placement for ccNUMA via hardware-generated memory traces

Non-uniform memory architectures with cache coherence (ccNUMA) are becoming increasingly common, not just for large-scale high performance platforms but also in the context of multi-core architectures. Under ccNUMA, data placement may influence overall application performance significantly as references resolved locally to a processor/core impose lower latencies than remote ones. This work develops a novel hardware-assisted page placement paradigm based on automated tracing of the memory references made by application threads. Two placement schemes, modeling both singlelevel and multi-level latencies, allocate pages near processors that most frequently access that memory page. These schemes leverage performance monitoring capabilities of contemporary microprocessors to efficiently extract an approximate trace of memory accesses. This information is used to decide page affinity, i.e., the node to which the page is bound. The method operates entirely in user space, is widely automated, and handles not only static but also dynamic memory allocation. Experiments show that this method, although based on lossy tracing, can efficiently and effectively improve page placement, leading to an average wall-clock execution time saving of over 20% for the tested benchmarks on the SGI Altix with a 2x remote access penalty and 12% on AMD Opterons with a 1.3–2.0x access penalty. This is accompanied by a one-time tracing overhead of 2.7% over the overall original program wallclock time.

Computer Science↗

A microbenchmark characterization of the Emu chick

The Emu Chick is a prototype system designed around the concept of migratory memory-side processing. Rather than transferring large amounts of data across power-hungry, high-latency interconnects, the Emu Chick moves lightweight thread contexts to near-memory cores before the beginning of each memory read. The current prototype hardware uses FPGAs to implement cache-less “Gossamer” cores for computational work and rely on a typical stationary core (PowerPC) to run basic operating system functions and migrate threads between nodes. In this multi-node characterization of the Emu Chick, we extend an earlier single-node investigation of the the memory bandwidth characteristics of the system through benchmarks like STREAM, pointer chasing, and sparse matrix-vector multiplication. Here, we compare the Emu Chick hardware to architectural simulation and an Intel Xeon-based platform. Our results demonstrate that for many basic operations the Emu Chick can use available memory bandwidth more efficiently than a more traditional, cache-based architecture although bandwidth usage suffers for computationally intensive workloads like SpMV. Moreover, the Emu Chick provides stable, predictable performance with up to 65% of the peak bandwidth utilization on a random-access pointer chasing benchmark with weak locality.

97 MATHEMATICS AND COMPUTING↗

Blackcomb: Hardware-Software Co-design for Non-Volatile Memory in Exascale Systems

Summary of technical results of Blackcomb Memory Devices We explored various different memory technologies (STTRAM, PCRAM, FeRAM, and ReRAM). The progress can be classified into three categories, below. Modeling and Tool Releases Various modeling tools have been developed over the last decade to help in the design of SRAM or DRAM-based memory hierarchies. To explore new design opportunities that NVM technologies can bring to the designers, we have developed similar high-level models for NVM, including PCRAMsim [Dong 2009], NVSim [Dong 2012], and NVMain [Poremba 2012]. NVSim is a circuit-level model for NVM performance, energy, and area estimation, which supports various NVM technologies, including STT-RAM, PCRAM, ReRAM, and legacy NAND Flash. NVSim is successfully validated against industrial NVM prototypes, and it is expected to help boost architecture-level NVM-related studies. On the other side, NVMain is a cycle accurate main memory simulator designed to simulate emerging nonvolatile memories at the architectural level. We have released these models as open source tools and provided contiguous support to them. We also proposed PS3-RAM, which is a fast, portable and scalable statistical STT-RAM reliability analysis model [Wen 2012]. Design Space Exploration and Optimization With the support of these models, we explore different device/circuit optimization techniques. For example, in [Niu 2012a] we studied the power reduction technique for the application of ECC scheme in ReRAM designs and proposed to use ECC code to relax the BER (Bit Error Rate) requirement of a single memory to improve the write energy consumption and latency for both 1T1R and cross-point ReRAM designs. In [Xu 2011], we proposed a methodology to design STT-RAM for different optimization goals such as read performance, write performance and write energy by leveraging the trade-off between write current and write time of MTJ. We also studied the tradeoffs in building a reliable crosspoint ReRAM array [Niu 2012b]. We have conducted an in depth analysis of the circuit and system level design implications of multi-layer cross-point Resistive RAM (MLCReRAM) from performance, power and reliability perspectives [Xu 2013]. The objective of this study is to understand the design trade-offs of this technology with respect to the MLC Phase Change Memory (MLCPCM).Our MLC ReRAM design at the circuit and system levels indicates that different resistance allocation schemes, programming strategies, peripheral designs, and material selections profoundly affect the area, latency, power, and reliability of MLC ReRAM. Based on this analysis, we conduct two case studies: first we compare MLC ReRAM design against MLC phase-change memory (PCM) and multi-layer cross-point ReRAM design, and point out why multi-level ReRAM is appealing; second we further explore the design space for MLC ReRAM. Architecture and Application We explored hybrid checkpointing using phase-change memory for future exascale systems [Dong 2011] and showed that the use of nonvolatile memory for local checkpointing significantly increases the number of faults covered by local checkpoints and reduces the probability of a global failure in the middle of a global checkpoint to less than 1%. We also proposed a technique called i2WAP to mitigate the write variations in NVM-based last-level cache for the improvement of the NVM lifetime [Wang 2013]. Our wear leveling technique attempts to work around the limitations of write endurance by arranging data access so that write operations can be distributed evenly across all the storage cells. During our intensive research on fault-tolerant NVM design, we found that ECC cannot effectively tolerate hard errors from limited write endurance and process imperfection. Therefore, we devised a novel Point and Discard (PAD) architecture in in [ 2012] as a hard-error-tolerant architecture for ReRAM-based Last Level Caches. PAD improves the lifetime of ReRAM caches by 1.6X-440X under different process variations without performance overhead in the system's early life. We have investigated the applicability of NVM for persistent memory design [Zhao 2013]. New byte addressable NVM enables fast persistent memory that allows in-memory persistent data objects to be updated with much higher throughput. Despite the significant improvement, the performance of these designs is only 50% of the native system with no persistence support, due to the logging or copy-on-write mechanisms used to update the persistent memory. A challenge in this approach is therefore how to efficiently enable atomic, consistent, and durable updates to ensure data persistence that survives application and/or system failures. We have designed a persistent memory system, called Klin, that can provide performance as close as that of the native system. The Klin design adopts a non-volatile cache and a non-volatile main memory for constructing a multi-versioned durable memory system, enabling atomic updates without logging or copy-on-write. Our evaluation shows that the proposed Kiln mechanism can achieve up to 2X of performance improvement to NVRAM-based persistent memory employing write-ahead logging. In addition, our design has numerous practical advantages: a simple and intuitive abstract interface, microarchitecture-level optimizations, fast recovery from failures, and no redundant writes to slow non-volatile storage media. The work was published in MICRO 2013 and received Best Paper Honorable Mentioned Award.

97 MATHEMATICS AND COMPUTING↗

Data Locality Enhancement of Dynamic Simulations for Exascale Computing (Final Report)

The development of modern processors exhibits two trends that complicate the optimizations of modern software. The first is the increasing sensitivity of processors' throughput to irregularities in computation. With more processors produced through a massive integration of simple cores, future systems will increasingly favor regular data-level parallel computations, but deviate from the needs of applications with complex patterns. Some evidences are already shown on Graphic Processing Units (GPU): Irregular data accesses (e.g., indirect references A[D[i]]) and conditional branches are limiting many GPU applications' performance at a level an order of magnitude lower than the peak of GPU. The second hardware trend is the growing gap between memory bandwidth and the aggregate speed—that is, the sum of all cores' computing power—of a Chip Multiprocessor (CMP). Despite the capped growth of the peak CPU speed, the aggregate speed of a CMP keeps increasing as more cores get into a single chip. It is expected that by 2018, node concurrency in an exascale system will increase by hundreds of times, whereas, memory bandwidth will expand by only 10 to 20 times. Consequently, data movement and storage is expected to consume more than 70% of the total system power. Bridging this gap is difficult; the complexities of modern CMP memory hierarchy make it even harder: Data cache becomes shared among computing units, and the sharing is often non-uniform—whether two computing units share a cache depends on their proximity and the level of the cache. On the recent IBM Power7 architecture, for instance, four hardware contexts (or SMT threads) in a core share the entire memory hierarchy, all cores in one chip share an on-chip L3 cache, and cores across chips share L3 and main memory through off-chip connections. These two trends complicate the translation of computing power into performance, especially for a program with either intensive data accesses or complex patterns in data accesses or control flow paths. Unfortunately, both attributes present and will persist in a class of important applications. For instance, many scientific simulations deal with a large volume of data. And meanwhile, as most real-world processes are non-uniform and evolving (e.g., the evolution of a galaxy or the process of a drug injection), both the computations and data accesses of these programs tend to be irregular and dynamically changing. Currently, the lack of support to these applications on modern CMP severely limits their performance. On GPU, as our recent study shows and other studies echo, performance enhancement of a factor of integers is possible when memory accesses or control flows are streamlined for a set of GPU applications. On multicore CPU, our studies show that traditional locality enhancement, for being oblivious to the new features of multicore memory hierarchy, may even cause large slowdown to data-intensive dynamic applications. The severity of the issues is expected to worsen as the two hardware trends continue. Some recent studies try to match software with the trends, but in a limited scope or manner. For irregularities on GPU, most studies focus on irregularities analyzable through static analysis (e.g., data accesses in regular loops). Dynamic irregularities are harder to address because the needed analysis and transformations typically have to happen at run time. Some other research resorts to hardware extensions, an actual adoption of which is unclear for the entailed space cost and complexity. For data locality, recent years have seen some exploitations of the new memory hierarchy on multicore for performance, but most of them are on process or thread scheduling, rather than program transformations. Our recent study reveals that program-level transformations may magnify the scheduling benefits by a factor of seven, concluding that program-level transformation should play a central role for data locality enhancement on modern CMP. But research in this direction has been sparse, and most have focused on data layout or cache performance modeling, rather than program transformations to match with the new memory hierarchy features. Overall, it is still an open question how to bridge the gap between dynamic computations and the two prominent properties of modern processors. The goal of this project is to develop a set of techniques and software tools to enhance the matching between memory accesses in dynamic simulations and the prominent features of modern and future CMP, alleviating the memory performance issues for petascale and exascale computing. This report summarizes the discoveries and products produced throughout this project. It includes free launch, a new software approach to overcoming the shortcomings of both methods; coherence-free multiview, an approach that allows multiple views of a single data object to co-exist on GPU memory during a GPU kernel execution; algorithmic optimizations to data analytics problems, especially those that involve lots of distance calculations. Supporting dynamic parallelism is important for GPU to benefit a broad range of applications. There are currently two fundamental ways for programs to exploit dynamic parallelism on GPU: a software-based approach with software-managed worklists, and a hardware-based approach through dynamic subkernel launches. Neither is satisfactory. The former is complicated to program and is often subject to some load imbalance; the latter suffers large runtime overhead. In this work, we propose free launch, a new software approach to overcoming the shortcomings of both methods. It allows programmers to use subkernel launches to express dynamic parallelism. It employs a novel compiler-based code transformation named subkernel launch removal to replace the subkernel launches with the reuse of parent threads. Coupled with an adaptive task assignment mechanism, the transformation reassigns the tasks in the subkernels to the parent threads with a good load balance. The technique requires no hardware extensions, immediately deployable on existing GPUs. It keeps the programming convenience of the subkernel launch-based approach while avoiding its large runtime overhead. Meanwhile, its superior load balancing makes it outperform manual worklist-based techniques by 3X on average. The work was published at the 48th Annual IEEE/ACM International Symposium on Microarchitecture (Micro'2015). A Graphic Processing Unit (GPU) system is typically equipped with many types of memory (e.g., global, constant, texture, shared, cache). Data placement determines what data are placed on which type of memory, essential for GPU memory performance. Prior optimizations of data placement always require a single view of a data object on memory, which limits the optimization effectiveness. In this work, we propose coherence-free multiview, an approach that allows multiple views of a single data object to co-exist on GPU memory during a GPU kernel execution. We demonstrate that under certain conditions, the multiple views can remain incoherent while facilitating enhanced data placement. We present a theorem and some compiler support to ensure the soundness of the usage of coherence-free multiview. We further develop reference-discerning data placement, a new way to enhance data placements on GPU. It enables more flexible data placements by using coherence-free multiview to leverage the slack in coherence requirement of some GPU programs. Experiments on three types of GPU systems show that, with less than 200KB space cost, the new data placement technique can provide a 1.6X average (up to 4.27X) speedup. In addition, we have examined algorithmic optimizations to data analytics problems, especially those that involve lots of distance calculations. Computing distances among data points is an essential part of many important algorithms in data analytics, graph analysis, and other domains. In each of these domains, developers have spent significant manual effort optimizing algorithms, often through novel applications of the triangle equality, in order to minimize the number of distance computations in the algorithms. In this work, we observe that many algorithms across these domains can be generalized as an instance of a generic distance-related abstraction. Based on this abstraction, we derive seven principles for correctly applying the triangular inequality to optimize distance-related algorithms. Guided by the findings, we develop Triangular OPtimizer (TOP), the first software framework that is able to automatically produce optimized algorithms that either matches or outperforms manually designed algorithms for solving distance-related problems. TOP achieves up to 237x speedups and 2.5X on average. The work has been published at the 32nd International Conference on Machine Learning (ICML'15) and the 41st International Conference on Very Large Data Bases (VLDB'15).

42 ENGINEERING↗

Virtual machine provisioning, code management, and data movement design for the Fermilab HEPCloud Facility

The Fermilab HEPCloud Facility Project has as its goal to extend the current Fermilab facility interface to provide transparent access to disparate resources including commercial and community clouds, grid federations, and HPC centers. This facility enables experiments to perform the full spectrum of computing tasks, including data-intensive simulation and reconstruction. We have evaluated the use of the commercial cloud to provide elasticity to respond to peaks of demand without overprovisioning local resources. Full scale data-intensive workflows have been successfully completed on Amazon Web Services for two High Energy Physics Experiments, CMS and NOνA, at the scale of 58000 simultaneous cores. This paper describes the significant improvements that were made to the virtual machine provisioning system, code caching system, and data movement system to accomplish this work. The virtual image provisioning and contextualization service was extended to multiple AWS regions, and to support experiment-specific data configurations. A prototype Decision Engine was written to determine the optimal availability zone and instance type to run on, minimizing cost and job interruptions. We have deployed a scalable on-demand caching service to deliver code and database information to jobs running on the commercial cloud. It uses the frontiersquid server and CERN VM File System (CVMFS) clients on EC2 instances and utilizes various services provided by AWS to build the infrastructure (stack). We discuss the architecture and load testing benchmarks on the squid servers. We also describe various approaches that were evaluated to transport experimental data to and from the cloud, and the optimal solutions that were used for the bulk of the data transport. Finally, we summarize lessons learned from this scale test, and our future plans to expand and improve the Fermilab HEP Cloud Facility.

72 PHYSICS OF ELEMENTARY PARTICLES AND FIELDS↗

Rapid Memory Footprint Access Diagnostics

Footprint and reuse distance measure temporal locality and therefore do not capture the significance of access patterns (spacial locality). A strided access pattern has the largest possible footprint but usually has the best performance. To highlight exposed memory latency, we separate footprint into strided (prefetchable) and irregular (non-prefetchable) access components and calculate the growth rate of each. To rapidly compute these footprint access diagnostics, we present two methods, whole-program and precise. Current footprint analyses can cause 200× or more slowdown with realistic inputs and are therefore impractical. Our whole-program method reduces the overhead to 10% by computing upper bounds, but still yields inter-procedural insight through a call path profile. Our precise method uses additional static analysis and profiling to refine the upper bounds for intra-procedural loop nests. We evaluate our approaches using benchmarks that vary access patterns (strided \vs unpredictable), sparsity (all words in a cache line \vs some), and reuse (varying and repeated accesses per element). Notably, for loop nests with unpredictable accesses, the precise method's accuracy is within 10% of ideal. The whole-program method has sufficient accuracy to diagnose bottlenecks.

data locality, memory footprint, footprint access ↗

Roofline Analysis in the Intel® Advisor to Deliver Optimized Performance for applications on Intel® Xeon Phi™ Processor

In this session we show, in two case studies, how the roofline feature of Intel Advisor has been utilized to optimize the performance of kernels of the XGC1 and PICSAR codes in preparation for Intel Knights Landing architecture. The impact of the implemented optimizations and the benefits of using the automatic roofline feature of Intel Advisor to study performance of large applications will be presented. This demonstrates an effective optimization strategy that has enabled these science applications to achieve up to 4.6 times speed-up and prepare for future exascale architectures. # Goal/Relevance of Session The roofline model [1,2] is a powerful tool for analyzing the performance of applications with respect to the theoretical peak achievable on a given computer architecture. It allows one to graphically represent the performance of an application in terms of operational intensity, i.e. the ratio of flops performed and bytes moved from memory in order to guide optimization efforts. Given the scale and complexity of modern science applications, it can often be a tedious task for the user to perform the analysis on the level of functions or loops to identify where performance gains can be made. With new Intel tools, it is now possible to automate this task, as well as base the estimates of peak performance on measurements rather than vendor specifications. The goal of this session is to demonstrate how the roofline feature of Intel Advisor can be used to balance memory vs. computation related optimization efforts and effectively identify performance bottlenecks. A series of typical optimization techniques: cache blocking, structure refactoring, data alignment, and vectorization illustrated by the kernel cases will be addressed. # Description of the codes ## XGC1 The XGC1 code [3] is a magnetic fusion Particle-In-Cell code that uses an unstructured mesh for its Poisson solver that allows it to accurately resolve the edge plasma of a magnetic fusion device. After recent optimizations to its collision kernel [4], most of the computing time is spent in the electron push (pushe) kernel, where these optimization efforts have been focused. The kernel code scaled well with MPI+OpenMP but had almost no automatic compiler vectorization, in part due to indirect memory addresses and in part due to low trip counts of low-level loops that would be candidates for vectorization. Particle blocking and sorting have been implemented to increase trip counts of low-level loops and improve memory locality, and OpenMP directives have been added to vectorize compute-intensive loops that were identified by Advisor. The optimizations have improved the performance of the pushe kernel 2x on Haswell processors and 1.7x on KNL. The KNL node-for-node performance has been brought to within 30% of a NERSC Cori phase I Haswell node and we expect to bridge this gap by reducing the memory footprint of compute intensive routines to improve cache reuse. ## PICSAR is a Fortran/Python high-performance Particle-In-Cell library targeting at MIC architectures first designed to be coupled with the PIC code WARP for the simulation of laser-matter interaction and particle accelerators. PICSAR also contains a FORTRAN stand-alone kernel for performance studies and benchmarks. A MPI domain decomposition is used between NUMA domains and a tile decomposition (cache-blocking) handled by OpenMP has been added for shared-memory parallelism and better cache management. The so-called current deposition and field gathering steps that compose the PIC time loop constitute major hotspots that have been rewritten to enable more efficient vectorization. Particle communications between tiles and MPI domain has been merged and parallelized. All considered, these improvements provide speedups of 3.1 for order 1 and 4.6 for order 3 interpolation shape factors on KNL configured in SNC4 quadrant flat mode. Performance is similar between a node of cori phase 1 and KNL at order 1 and better on KNL by a factor 1.6 at order 3 with the considered test case (homogeneous thermal plasma).

97 MATHEMATICS AND COMPUTING↗