Search NASA⌕ Search

SEARCH · Search NASA

Results for “memory access optimization”

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 217 records · Page 12

Reverse-Mode Algorithmic Differentiation of an OpenMP-Parallel Compressible Flow Solver

y Reverse-mode algorithmic differentiation (AD) is an established method for obtaining adjoint derivatives of computer simulation applications. In computational fluid dynamics (CFD), adjoint derivatives of a cost function output such as drag or lift with respect to design parameters such as surface coordinates or geometry control points are a key ingredient for shape optimization, uncertainty quantification and flow control. The computational cost of CFD applications and their derivatives makes it essential to use high-performance computing hardware efficiently, including multi- and many-core architectures. Nevertheless, OpenMP is not supported in most AD tools, and previously shown methods achieve poor scalability of the derivative code. We present the AD of an OpenMP-parallelized finite volume compressible flow solver for unstructured meshes. Our approach enables us to reuse the parallelization of the original code in the computation of adjoint derivatives. The method works by identifying code segments that can be differentiated in reverse-mode without changing their memory access pattern. The OpenMP parallelization is integrated into the derivative code during the build process in a way that is robust to modifications of the original code and independent of the OpenMP support of the differentiation tool. We show the scalability of our adjoint CFD solver on test cases ranging from thousands to millions of finite volume mesh cells on CPUs with up to 16 threads as well as on an Intel XeonPhi card with 236 threads. We demonstrate that our approach is more practical to implement for production-sized CFD codes and produces more efficient adjoint derivative code than previously shown AD methods.

Huckelheim, Jan↗

Achieving High Performance With TCP Over 40 GbE on NUMA Architectures for CMS Data Acquisition

TCP and the socket abstraction have barely changed over the last two decades, but at the network layer there has been a giant leap from a few megabits to 100 gigabits in bandwidth. At the same time, CPU architectures have evolved into the multicore era and applications are expected to make full use of all available resources. Applications in the data acquisition domain based on the standard socket library running in a Non-Uniform Memory Access (NUMA) architecture are unable to reach full efficiency and scalability without the software being adequately aware about the IRQ (Interrupt Request), CPU and memory affinities. During the first long shutdown of LHC, the CMS DAQ system is going to be upgraded for operation from 2015 onwards and a new software component has been designed and developed in the CMS online framework for transferring data with sockets. This software attempts to wrap the low-level socket library to ease higher-level programming with an API based on an asynchronous event driven model similar to the DAT uDAPL API. It is an event-based application with NUMA optimizations, that allows for a high throughput of data across a large distributed system. This paper describes the architecture, the technologies involved and the performance measurements of the software in the context of the CMS distributed event building.

46 INSTRUMENTATION RELATED TO NUCLEAR SCIENCE AND ↗

Distributed caching for processing raw arrays

As applications continue to generate multi-dimensional data at exponentially increasing rates, fast analytics to extract meaningful results is becoming extremely important. The database community has developed array databases that alleviate this problem through a series of techniques. In-situ mechanisms provide direct access to raw data in the original format---without loading and partitioning. Parallel processing scales to the largest datasets. In-memory caching reduces latency when the same data are accessed across a workload of queries. However, we are not aware of any work on distributed caching of multi-dimensional raw arrays. In this paper, we introduce a distributed framework for cost-based caching of multi-dimensional arrays in native format. Given a set of files that contain portions of an array and an online query workload, the framework computes an effective caching plan in two stages. First, the plan identifies the cells to be cached locally from each of the input files by continuously refining an evolving R-tree index. In the second stage, an optimal assignment of cells to nodes that collocates dependent cells in order to minimize the overall data transfer is determined. We design cache eviction and placement heuristic algorithms that consider the historical query workload. A thorough experimental evaluation over two real datasets in three file formats confirms the superiority -- by as much as two orders of magnitude -- of the proposed framework over existing techniques in terms of cache overhead and workload execution time.

Zhao, Weijie↗

High Level Synthesis of RDF Queries for Graph Analytics

In this paper we present a set of techniques that enable the synthesis of efficient custom accelerators for memory intensive, irregular applications. To address irregular applications challenges (large memory footprints, unpredictable fine- grained data accesses, and high synchronization intensity), and exploit their opportunities (thread level parallelism, memory level parallelism), we propose a novel accelerator design which take advantage of an adaptive and Distributed Controller (DC) architecture, and a Memory Interface (MI) that supports parallel memory subsystems. Among the multitude of algorithms that may benefit from our solution, we focus on the acceleration of graph analytics applications, and in particular, on the syn- thesis of SPARQL queries on Resource Description Framework (RDF) databases. We achieve this objective by incorporating the synthesis techniques into Bambu, an Open Source high- level synthesis tools, and interfacing the system with GEMS, the Graph database Engine for Multithreaded Systems. The front- end of GEMS generates optimized C implementations of the input queries, modeled as pattern matching routines, which are then automatically synthesized by Bambu. We validate our approach synthesizing several SPARQL queries from the Lehigh University Benchmark (LUBM).

HLS, FPGA, Convey, RDF, GEMS, Memory-intensive,↗

SU-E-T-531: Performance Evaluation of Multithreaded Geant4 for Proton Therapy Dose Calculations in a High Performance Computing Facility

Purpose: To evaluate the efficiency of multithreaded Geant4 (Geant4-MT, version 10.0) for proton Monte Carlo dose calculations using a high performance computing facility. Methods: Geant4-MT was used to calculate 3D dose distributions in 1×1×1 mm3 voxels in a water phantom and patient's head with a 150 MeV proton beam covering approximately 5×5 cm2 in the water phantom. Three timestamps were measured on the fly to separately analyze the required time for initialization (which cannot be parallelized), processing time of individual threads, and completion time. Scalability of averaged processing time per thread was calculated as a function of thread number (1, 100, 150, and 200) for both 1M and 50 M histories. The total memory usage was recorded. Results: Simulations with 50 M histories were fastest with 100 threads, taking approximately 1.3 hours and 6 hours for the water phantom and the CT data, respectively with better than 1.0 % statistical uncertainty. The calculations show 1/N scalability in the event loops for both cases. The gains from parallel calculations started to decrease with 150 threads. The memory usage increases linearly with number of threads. No critical failures were observed during the simulations. Conclusion: Multithreading in Geant4-MT decreased simulation time in proton dose distribution calculations by a factor of 64 and 54 at a near optimal 100 threads for water phantom and patient's data respectively. Further simulations will be done to determine the efficiency at the optimal thread number. Considering the trend of computer architecture development, utilizing Geant4-MT for radiotherapy simulations is an excellent cost-effective alternative for a distributed batch queuing system. However, because the scalability depends highly on simulation details, i.e., the ratio of the processing time of one event versus waiting time to access for the shared event queue, a performance evaluation as described is recommended.

60 APPLIED LIFE SCIENCES↗

Final Project Report: Data Locality Enhancement of Dynamic Simulations for Exascale Computing

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 manycore systems, alleviating the memory performance issues for exascale computing. In the first three years, the PI and his group have achieves some significant progress towards the goal, producing a set of novel techniques for improving the memory performance and data locality in manycore systems, yielding 18 conference and workshop papers and 4 journal papers and graduating 6 Ph.Ds. This report summarizes the research results of this project through that period.

97 MATHEMATICS AND COMPUTING↗

Software Coherence in Multiprocessor Memory Systems

Processors are becoming faster and multiprocessor memory interconnection systems are not keeping up. Therefore, it is necessary to have threads and the memory they access as near one another as possible. Typically, this involves putting memory or caches with the processors, which gives rise to the problem of coherence: if one processor writes an address, any other processor reading that address must see the new value. This coherence can be maintained by the hardware or with software intervention. Systems of both types have been built in the past; the hardware-based systems tended to outperform the software ones. However, the ratio of processor to interconnect speed is now so high that the extra overhead of the software systems may no longer be significant. This issue is explored both by implementing a software maintained system and by introducing and using the technique of offline optimal analysis of memory reference traces. It finds that in properly built systems, software maintained coherence can perform comparably to or even better than hardware maintained coherence. The architectural features necessary for efficient software coherence to be profitable include a small page size, a fast trap mechanism, and the ability to execute instructions while remote memory references are outstanding.

Bolosky, William Joseph↗

GPU Data Access on Complex Geometries for D3Q19 Lattice Boltzmann Method

GPU performance of the lattice Boltzmann method (LBM) depends heavily on memory access patterns. When LBM is advanced with GPUs on complex computational domains, geometric data is typically accessed indirectly, and lattice data is typically accessed lexicographically in the Structure of Array (SoA) layout. Although there are a variety of existing access patterns beyond the typical choices, no study has yet examined the relative efficacy between them. Here, we compare a suite of memory access schemes via empirical testing and performance modeling. We find strong evidence that semi-direct addressing is the superior addressing scheme for the majority of cases examined: Semi-direct addressing increases computational speed and often reduces memory consumption. For lattice layout, we find that the Collected Structure of Arrays (CSoA) layout outperforms the SoA layout. When compared to state-of-the-art practices, our recommended addressing modifications lead to performance gains between 10-40% across different complex geometries, fluid volume fractions, and resolutions. The modifications also lead to a decrease in memory consumption by as much as 17%. Having discovered these improvements, we examine a highly resolved arterial geometry on a leadership class system. On this system we present the first near-optimal strong results for LBM with arterial geometries run on GPUs. We also demonstrate that the above recommendations remain valid for large scale, many device simulations, which leads to an increased computational speed and average memory usage reductions. To understand these observations, we employ performance modeling which reveals that semi-direct methods outperform indirect methods due to a reduced number of total loads/stores in memory, and that CSoA outperforms SoA and bundling due to improved caching behavior.

Herschlag, Gregory↗

ACDT: Architected Composite Data Types Trading-in Unfettered Data Access for Improved Execution

With Exascale performance and its challenges in mind, one ubiquitous concern among architects is energy efficiency. Petascale systems projected to Exascale systems are unsustainable at current power consumption rates. One major contributor to system-wide power consumption is the number of memory operations leading to data movement and management techniques applied by the runtime system. To address this problem, we present the concept of the Architected Composite Data Types (ACDT) framework. The framework is made aware of data composites, assigning them a specific layout, transformations and operators. Data manipulation overhead is amortized over a larger number of elements and program performance and power efficiency can be significantly improved. We developed the fundamentals of an ACDT framework on a massively multithreaded adaptive runtime system geared towards Exascale clusters. Showcasing the capability of ACDT, we exercised the framework with two representative processing kernels - Matrix Vector Multiply and the Cholesky Decomposition – applied to sparse matrices. As transformation modules, we applied optimized compress/decompress engines and configured invariant operators for maximum energy/performance efficiency. Additionally, we explored two different approaches based on transformation opaqueness in relation to the application. Under the first approach, the application is agnostic to compression and decompression activity. Such approach entails minimal changes to the original application code, but leaves out potential applicationspecific optimizations. The second approach exposes the decompression process to the application, hereby exposing optimization opportunities that can only be exploited with application knowledge. The experimental results show that the two approaches have their strengths in HW and SW respectively, where the SW approach can yield performance and power improvements that are an order of magnitude better than ACDT-oblivious, hand-optimized implementations.We consider the ACDT runtime framework an important component of compute nodes that will lead towards power efficient Exascale clusters.

exascale↗

Analysis of GPU Data Access Patterns on Complex Geometries for the D3Q19 Lattice Boltzmann Algorithm

GPU performance of the lattice Boltzmann method (LBM) depends heavily on memory access patterns. When implemented with GPUs on complex domains, typically, geometric data is accessed indirectly and lattice data is accessed lexicographically. Although there are a variety of other options, no study has examined the relative efficacy between them. Here, we examine a suite of memory access schemes via empirical testing and performance modeling. We find strong evidence that semi-direct is often better suited than the more common indirect addressing, providing increased computational speed and reducing memory consumption. For the layout, we find that the Collected Structure of Arrays (CSoA) and bundling layouts outperform the common Structure of Array layout; on V100 and P100 devices, CSoA consistently outperforms bundling, however the relationship is more complicated on K40 devices. When compared to state-of-the-art practices, our recommendations lead to speedups of 10–40 percent and reduce memory consumption up to 17 percent. Using performance modeling and computational experimentation, we determine the mechanisms behind the accelerations. We demonstrate that our results hold across multiple GPUs on two leadership class systems, and present the first near-optimal strong results for LBM with arterial geometries run on GPUs.

42 ENGINEERING↗

Active Flash: Performance-Energy Tradeoffs for Out-of-Core Processing on Non-Volatile Memory Devices

In this abstract, we study the performance and energy tradeoffs involved in migrating data analysis into the flash device, a process we refer to as Active Flash. The Active Flash paradigm is similar to 'active disks', which has received considerable attention. Active Flash allows us to move processing closer to data, thereby minimizing data movement costs and reducing power consumption. It enables true out-of-core computation. The conventional definition of out-of-core solvers refers to an approach to process data that is too large to fit in the main memory and, consequently, requires access to disk. However, in Active Flash, processing outside the host CPU literally frees the core and achieves real 'out-of-core' analysis. Moving analysis to data has long been desirable, not just at this level, but at all levels of the system hierarchy. However, this requires a detailed study on the tradeoffs involved in achieving analysis turnaround under an acceptable energy envelope. To this end, we first need to evaluate if there is enough computing power on the flash device to warrant such an exploration. Flash processors require decent computing power to run the internal logic pertaining to the Flash Translation Layer (FTL), which is responsible for operations such as address translation, garbage collection (GC) and wear-leveling. Modern SSDs are composed of multiple packages and several flash chips within a package. The packages are connected using multiple I/O channels to offer high I/O bandwidth. SSD computing power is also expected to be high enough to exploit such inherent internal parallelism within the drive to increase the bandwidth and to handle fast I/O requests. More recently, SSD devices are being equipped with powerful processing units and are even embedded with multicore CPUs (e.g. ARM Cortex-A9 embedded processor is advertised to reach 2GHz frequency and deliver 5000 DMIPS; OCZ RevoDrive X2 SSD has 4 SandForce controllers, each with 780MHz max frequency Tensilica core). Efforts that take advantage of the available computing cycles on the processors on SSDs to run auxiliary tasks other than actual I/O requests are beginning to emerge. Kim et al. investigate database scan operations in the context of processing on the SSDs, and propose dedicated hardware logic to speed up scans. Also, cluster architectures have been explored, which consist of low-power embedded CPUs coupled with small local flash to achieve fast, parallel access to data. Processor utilization on SSD is highly dependent on workloads and, therefore, they can be idle during periods with no I/O accesses. We propose to use the available processing capability on the SSD to run tasks that can be offloaded from the host. This paper makes the following contributions: (1) We have investigated Active Flash and its potential to optimize the total energy cost, including power consumption on the host and the flash device; (2) We have developed analytical models to analyze the performance-energy tradeoffs for Active Flash, by treating the SSD as a blackbox, this is particularly valuable due to the proprietary nature of the SSD internal hardware; and (3) We have enhanced a well-known SSD simulator (from MSR) to implement 'on-the-fly' data compression using Active Flash. Our results provide a window into striking a balance between energy consumption and application performance.

42 ENGINEERING↗

ELSI: A unified software interface for Kohn–Sham electronic structure solvers

Solving the electronic structure from a generalized or standard eigenproblem is often the bottleneck in large scale calculations based on Kohn-Sham density-functional theory. This problem must be addressed by essentially all current electronic structure codes, based on similar matrix expressions, and by high-performance computation. We here present a unified software interface, ELSI, to access different strategies that address the Kohn-Sham eigenvalue problem. Currently supported algorithms include the dense generalized eigensolver library ELPA, the orbital minimization method implemented in libOMM, and the pole expansion and selected inversion (PEXSI) approach with lower computational complexity for semilocal density functionals. The ELSI interface aims to simplify the implementation and optimal use of the different strategies, by offering (a) a unified software framework designed for the electronic structure solvers in Kohn-Sham density-functional theory; (b) reasonable default parameters for a chosen solver; (c) automatic conversion between input and internal working matrix formats, and in the future (d) recommendation of the optimal solver depending on the specific problem. As a result, comparative benchmarks are shown for system sizes up to 11,520 atoms (172,800 basis functions) on distributed memory supercomputing architectures.

71 CLASSICAL AND QUANTUM MECHANICS, GENERAL PHYSIC↗

Distributed Halide

Many image processing tasks are naturally expressed as a pipeline of small computational kernels known as stencils. Halide is a popular domain-specific language and compiler designed to implement image processing algorithms. Halide uses simple language constructs to express what to compute and a separate scheduling co-language for expressing when and where to perform the computation. This approach has demonstrated performance comparable to or better than hand-optimized code. Until now, however, Halide has been restricted to parallel shared memory execution, limiting its performance for memory-bandwidth-bound pipelines or large-scale image processing tasks. We present an extension to Halide to support distributed-memory parallel execution of complex stencil pipelines. These extensions compose with the existing scheduling constructs in Halide, allowing expression of complex computation and communication strategies. Existing Halide applications can be distributed with minimal changes, allowing programmers to explore the tradeoff between recomputation and communication with little effort. Approximately 10 new of lines code are needed even for a 200 line, 99 stage application. On nine image processing benchmarks, our extensions give up to a 1.4× speedup on a single node over regular multithreaded execution with the same number of cores, by mitigating the effects of non-uniform memory access. The distributed benchmarks achieve up to 18× speedup on a 16 node testing machine and up to 57× speedup on 64 nodes of the NERSC Cori supercomputer.

97 MATHEMATICS AND COMPUTING↗

Memory Optimizations for Sparse Linear Algebra on GPU Hardware

An effort to maximize memory bandwidth utilization for a sparse linear algebra kernel executing on NVIDIA® Tesla V100 and A100 Graphics Processing Units (GPUs) is described. The kernel consists of a block-sparse matrix-vector product and a series of forward/backward triangular solves. The computation is memory-bound and exhibits low arithmetic intensity. Along with a relatively small block size, the data layout poses a challenge to effectively utilize the available memory bandwidth on common GPU architectures. An earlier implementation using a warp to process a single row of the matrix was found to yield good memory performance on the V100 architecture. However, anew approach, which assigns a warp to six rows of the matrix, is proposed for the A100. In addition, two new features offered by the A100 architecture are explored.L2residency control enables a portion of theL2cache to be used for persistent data access, and the asynchronous copy instruction allows data to be loaded directly from main memory into shared memory. Demonstrations show that the new implementation improves memory bandwidth utilization from 71.5% to 81.2% of the peak available on theA100 architecture.

GPU↗

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↗

Enabling Scalable and Extensible Memory-mapped Datastores in Userspace

Exascale workloads are expected to incorporate data-intensive processing in close coordination with traditional physics simulations. These emerging scientific, data-analytics and machine learning applications need to access a wide variety of datastores in flat files and structured databases. Programmer productivity is greatly enhanced by mapping datastores into the application process's virtual memory space to provide a unified “in-memory” interface. Currently, memory mapping is provided by system software primarily designed for generality and reliability. However, scalability at high concurrency is a formidable challenge on exascale systems. Also, there is a need for extensibility to support new datastores potentially requiring HPC data transfer services. In this article, we present UMap , a scalable and extensible userspace service for memory-mapping datastores. Furthermore, through decoupled queue management, concurrency aware adaptation, and dynamic load balancing, UMap enables application performance to scale even at high concurrency. We evaluate UMap in data-intensive applications, including sorting, graph traversal, database operations, and metagenomic analytics. Our results show that UMap as a userspace service outperforms an optimized kernel-based service across a wide range of intra-node concurrency by 1.22-1.9 × . We performed two case studies to demonstrate UMap 's extensibility. First, a new datastore residing in remote memory is incorporated into UMap as an application-specific plugin. Second, we present a persistent memory allocator Metall built atop UMap for unified storage/memory.

97 MATHEMATICS AND COMPUTING↗

Tag-Split Cache for Efficient GPGPU Cache Utilization

Modern GPUs employ cache to improve memory system efficiency. However, large amount of cache space is underutilized due to irregular memory accesses and poor spatial locality which exhibited commonly in GPU applications. Our experiments show that using smaller cache lines could improve cache space utilization, but it also frequently suffers from significant performance loss by introducing large amount of extra cache requests. In this work, we propose a novel cache design named tag-split cache (TSC) that enables fine-grained cache storage to address the problem of cache space underutilization while keeping memory request number unchanged. TSC divides tag into two parts to reduce storage overhead, and it supports multiple cache line replacement in one cycle.

performance and energy optimization↗

Multi-GPU implementation of a VMAT treatment plan optimization algorithm

Purpose: Volumetric modulated arc therapy (VMAT) optimization is a computationally challenging problem due to its large data size, high degrees of freedom, and many hardware constraints. High-performance graphics processing units (GPUs) have been used to speed up the computations. However, GPU’s relatively small memory size cannot handle cases with a large dose-deposition coefficient (DDC) matrix in cases of, e.g., those with a large target size, multiple targets, multiple arcs, and/or small beamlet size. The main purpose of this paper is to report an implementation of a column-generation-based VMAT algorithm, previously developed in the authors’ group, on a multi-GPU platform to solve the memory limitation problem. While the column-generation-based VMAT algorithm has been previously developed, the GPU implementation details have not been reported. Hence, another purpose is to present detailed techniques employed for GPU implementation. The authors also would like to utilize this particular problem as an example problem to study the feasibility of using a multi-GPU platform to solve large-scale problems in medical physics. Methods: The column-generation approach generates VMAT apertures sequentially by solving a pricing problem (PP) and a master problem (MP) iteratively. In the authors’ method, the sparse DDC matrix is first stored on a CPU in coordinate list format (COO). On the GPU side, this matrix is split into four submatrices according to beam angles, which are stored on four GPUs in compressed sparse row format. Computation of beamlet price, the first step in PP, is accomplished using multi-GPUs. A fast inter-GPU data transfer scheme is accomplished using peer-to-peer access. The remaining steps of PP and MP problems are implemented on CPU or a single GPU due to their modest problem scale and computational loads. Barzilai and Borwein algorithm with a subspace step scheme is adopted here to solve the MP problem. A head and neck (H and N) cancer case is then used to validate the authors’ method. The authors also compare their multi-GPU implementation with three different single GPU implementation strategies, i.e., truncating DDC matrix (S1), repeatedly transferring DDC matrix between CPU and GPU (S2), and porting computations involving DDC matrix to CPU (S3), in terms of both plan quality and computational efficiency. Two more H and N patient cases and three prostate cases are used to demonstrate the advantages of the authors’ method. Results: The authors’ multi-GPU implementation can finish the optimization process within ∼1 min for the H and N patient case. S1 leads to an inferior plan quality although its total time was 10 s shorter than the multi-GPU implementation due to the reduced matrix size. S2 and S3 yield the same plan quality as the multi-GPU implementation but take ∼4 and ∼6 min, respectively. High computational efficiency was consistently achieved for the other five patient cases tested, with VMAT plans of clinically acceptable quality obtained within 23–46 s. Conversely, to obtain clinically comparable or acceptable plans for all six of these VMAT cases that the authors have tested in this paper, the optimization time needed in a commercial TPS system on CPU was found to be in an order of several minutes. Conclusions: The results demonstrate that the multi-GPU implementation of the authors’ column-generation-based VMAT optimization can handle the large-scale VMAT optimization problem efficiently without sacrificing plan quality. The authors’ study may serve as an example to shed some light on other large-scale medical physics problems that require multi-GPU techniques.

60 APPLIED LIFE SCIENCES↗