Multidimensional Grid Aware Address Prediction for GPGPU

Graphics Processing Units (GPUs) contain multiple SIMD cores and each core can run a large number of threads concurrently. Threads in a core are scheduled and executed infixed sized groups, called warps. Each core contains one or more warp schedulers that select and execute warps from a pool of ready warps. In spite of having a large number of concurrent warps - 48 on NVIDIA Fermi architecture GPU - on manyGPGPU applications, current warp scheduling algorithms can not effectively utilize the hardware resources, resulting in stall cycles and loss in performance. The main reason for this is current warp scheduling algorithms mostly focus on long latency operations, especially global memory accesses, and do not take into account factors such as the progress of each thread block and the number of ready warps. In this paper, we propose, PRO, a progress warp scheduling algorithm that not only focuses on finishing individual thread blocks faster but also on reducing the overall execution time. These goals are achieved by dynamically prioritizing thread blocks and warps, based on their progress. We implemented our proposed algorithm in the GPGPU-SIM simulator and evaluated on various applications from GPGPU-SIM, Rosina and CUDASDK benchmark suites. We achieved an average speedup of 1.12x and a maximum speedup of 1.94x over the commonly used LooseRound Robin warp scheduling algorithm. Over the Two Level warp scheduler, our algorithm showed an average speedup of1.13x and a maximum speedup of 1.6x. Our proposed solution requires only a very small increase in the GPU hardware.

The ability to perform fast context-switching and massive multi-threading is the forte of modern GPU architectures, which have emerged as an efficient alternative to traditional chip-multiprocessors for parallel workloads. One of the main benefits of such architecture is its latency-hiding capability. However, the efficacy of GPU's latency-hiding varies significantly across GPGPU applications. To investigate this, this paper first proposes a new algorithm that profiles execution behavior of GPGPU applications. We characterize latencies caused by various pipeline hazards, memory accesses, synchronization primitives, and the warp scheduler. Our results show that the current round-robin warp scheduler works well in overlapping various latency stalls with the execution of other available warps for only a few GPGPU applications. For other applications, there is an excessive latency stall that cannot be hidden by the scheduler effectively. With the latency characterization insight, we observe a significant execution time disparity for warps within the same thread block, which causes sub-optimal performance, called the warp criticality problem. To tackle the warp criticality problem, we design a family of criticality-aware warp scheduling (CAWS) policies by scheduling the critical warp(s) more frequently than other warps. Our results on the breadth-first-search, B+tree search, two point angular correlation function, and K-means clustering show that, with oracle knowledge of warp criticality, our best-performing scheduling policy can improve GPGPU applications' performance by 17% on average. With our designed criticality predictor, the various scheduling policies can improve performance by 10-21% on breadth-first-search. To our knowledge, this is the first paper to characterize warp criticality and explore different criticality-aware warp scheduling policies for GPGPU workloads.

In-memory Key-Value stores (IMKVs) provide significantly higher performance than traditional disk-based counterparts. As memory technologies advance, IMKVs become practical for modern Big Data processing, which include financial services, e-commerce, telecommunication network, etc. Recently, various IMKVs have been proposed from both academia and industrial. In order to leverage high performance random access capability of main memory, most IMKVs employ hashing based index structures to retrieve data according to keys. Consequently, a regular memory access pattern can be observed in data retrieval from those IMKVs. Normally speaking, one access to index (hash table), which is also located in main memory, is followed by another memory access to value data. Such a regular access pattern provides a potential opportunity that data prefetching techniques can be employed to improve memory access efficiency for data retrieval in these IMKVs. Based on this observation, we explore various data prefetching techniques with proper architecture level modifications on memory controller considering trade-off between design overhead and performance. Specifically, we focus on two key design issues of prefetching techniques: (1) where to fetch data (i.e. data address)? and (2) how many data to fetch (i.e. data size)? Experimental results demonstrate that memory access performance can be substantially improved up to 35.4 %. In addition, we also demonstrate the overhead of prefetching on power consumption.

Irregular data references are difficult to prefetch, as the future memory address of a load instruction is hard to anticipate by a compiler. However, recent studies as well as our experience indicate that some important load instructions in irregular programs contain stride access patterns. Although the load instructions with stride patterns are difficult to identify with static compiler techniques, we developed an efficient profiling method to discover these load instructions. The new profiling method integrates the profiling for stride information and the traditional profiling for edge frequency into a single profiling pass. The integrated profiling pass runs only 17% slower than the frequency profiling alone. The collected stride information helps the compiler to identify load instructions with stride patterns that can be prefetched efficiently and beneficially. We implemented the new profiling and prefetching techniques in a research compiler for Itanium Processor Family (IPF), and obtained significant performance improvement for the SPECINT2000 programs running on Itanium machines. For example, we achieved a 1.59x speedup for 181.mcf, 1.14x for 254.gap, and 1.08x for 197.parser. We also showed that the performance gain is stable across input data sets. These benefits make the new profiling and prefetching techniques suitable for production compilers.

Graphics processing units(GPUs) have been increasingly used to accelerate general purpose computations. By exploiting massive thread-level parallelism (TLP), GPUs can achieve high throughput as well as memory latency hiding. As a result, a very large register file (RF) is typically required to enable fast and low-cost context switching between tens of thousands of active threads. However, RF resource is still insufficient to enable all thread level parallelism and the lack of RF resources can hurt performance by limiting the occupancy of GPU threads. Moreover, if the available RF capacity can not fit the requirement of a thread block, GPU needs to fetch some variables from local memory which may lead to long memory access latencies. By observing that a large percentage of computed results actually have fewer significant bits compared to the full width of a 32-bit register for many GPGPU applications, we propose a GPU register packing scheme to dynamically exploit narrowwidth operands and pack multiple operands into a single fullwidth register. By using dynamically register packing, more RF space is available which allows GPU to enable more TLP through assigning additional thread blocks on SMs (Streaming Multiprocessors) and thus improve performance. The experimental results show that our GPU register packing scheme can achieve up to 1.96X speedup and 1.18X on average.

General-purpose workloads running on modern graphics processing units (GPGPUs) rely on hardware-based barriers to synchronize warps within a thread block (TB). However, imbalance may exist before reaching a barrier if a GPGPU workload contains irregular memory accesses, i.e., some warps may be critical while others may not. Ideally, cache space should be reserved for the critical warps. Unfortunately, current cache management policies are unaware of the existence of barriers and critical warps, which significantly limits the performance of irregular memory-intensive GPGPU workloads.In this work, we propose Barrier-Aware Cache Management (BACM), which is built on top of two underlying policies: a greedy policy and a friendly policy. The greedy policy does not allow non-critical warps to allocate cache lines in the L1 data cache; only critical warps can. The friendly policy allows non-critical warps to allocate cache lines but only over invalid or lower-priority cache lines. Based on the L1 data cache hit rate of non-critical warps, BACM dynamically chooses between the greedy and friendly policies. By doing so, BACM reserves more cache space to accelerate critical warps, thereby improving overall performance. Experimental results show that BACM achieves an average performance improvement of 24% and 20% compared to the GTO and BAWS policies, respectively. BACM's hardware cost is limited to 96 bytes per streaming multiprocessor.

Objective: To validate the 17-gene Oncotype DX Genomic Prostate Score (GPS) biopsy-based gene expression assay as a predictor of adverse pathology (AP: Pathologic Gleason score ≥4+3, presence of any Gleason 5, and/or ≥pT3) in African American (AA) men. Methods: Between February 2009 and September 2014, AA and European American (EA) men with very low, low, and intermediate risk prostate cancer (PCa) enrolled in a multi-institutional prospective study of vitamin D impacts of biopsy outcomes. The subset who proceeded to immediate radical prostatectomy (RP) after biopsy with available biopsy tumor blocks was included in a comparative effectiveness analysis of GPS on biopsy and its association with surgical AP on RP using logistic regression and receiver operating characteristic curves. Multiplicative interactions tested for differential prediction of GPS accuracy by race (AA vs. EA). Results: Overall, 102 AA and 76 EA men elected RP, out of which 51 (47.2%) had AP. GPS result was a significant predictor of AP (odds ratio per 20 GPS units [OR/20 units] in AA: 4.78; 95% CI 1.8-12.5, P =0.001; and EA: 4.41; 95% CI 1.6-11.9, P =0.003) in univariable analysis. On multivariate analysis, there was a significant interaction between GPS and race (P=0.01). On race stratified binary logistic regression, AP remained significant after adjustment for NCCN risk group in both AA and EA men (OR/20 units in AA: 3.29; 95% CI 1.2-9.1, P =0.02; and EA: 4.24; 95% CI 1.4-12.6, P =0.01). Area under the curves for AP using GPS/20 units was 0.719 for AAs vs. 0.745 for EAs (P=0.39). Conclusion: In this AA validation study, the Oncotype Dx PCa assay was confirmed as an independent predictor of AP at prostatectomy in AA and EA men with similar predictive accuracy, though there was evidence of effect modification by race. Citation Format: Samuel Carbunaru, Virgilia Marcias, Peter Gann, Roohollah Sharifi, Ximing Yang, Michael Dixon, Chase Gorbein, Borko Jovanovic, Andre Kajdacsy-Balla, Adam B. Murphy. Oncotype DX assay has similar predictive accuracy for adverse pathology at radical prostatectomy in African American and European American men [abstract]. In: Proceedings of the Twelfth AACR Conference on the Science of Cancer Health Disparities in Racial/Ethnic Minorities and the Medically Underserved; 2019 Sep 20-23; San Francisco, CA. Philadelphia (PA): AACR; Cancer Epidemiol Biomarkers Prev 2020;29(6 Suppl_2):Abstract nr A067.

Cancer is one of the leading causes of mortality in the developed world, and prognostic assessment of cancer patients is indispensable in medical care. Medical researchers are accustomed to using regression models to predict patient outcomes. Neural networks have been proposed as an alternative with great potential. Nonetheless, empirical evidence remains lacking to support the application of this technique as the appropriate method to investigate cancer prognosis. Utilizing data on patients from two National Cancer Institute of Canada clinical trials, we compared predictive accuracy of neural network models and logistic regression models on risk of death of limited-stage small-cell lung cancer patients. Our results suggest that neural network and logistic regression models have similar predictive accuracy. The distributions of individual predicted probabilities are very similar. On occasion, however, the prediction pairs are quite different, suggesting that they do not always give the same interpretations of the same variables.

515 Background: Survival outcomes are heterogeneous across different subsets of mRCC pts. Poor risk pts have grim outcomes and are less likely to be enrolled in clinical trials, hence the lack of prospective data characterizing this pt population. We provide a comprehensive analysis of poor risk pts, as defined by three widely used prognostic models, in the era of targeted therapy. Methods: We conducted a pooled retrospective analysis of 4,736 mRCC pts treated on phase II and III clinical trials. Pts were defined as poor risk according to the Memorial Sloan Kettering Cancer Center (MSKCC), International Metastatic Database Consortium (IMDC), and Hudes prognostic risk models. We evaluated overall survival (OS), progression-free survival (PFS), objective response rate (ORR), and adverse events (AEs) in poor risk pts in each of the risk models in the total cohort and in a subset of pts who remained on treatment for >12 months (mos). The models’ concordance (c-) index was used to assess their prognostic performance. Results: We identified 1145 (24.2%), 904 (19.1%), and 1901 (40.1%) poor risk pts by IMDC, MSKCC, and Hudes models, respectively. The majority were <65 years of age (70-73%) with clear-cell histology (85-88%), prior nephrectomy (62-66%), and no prior systemic therapy (80-86%). The efficacy and toxicity outcomes as well as the c-indexes for each of the models are reported in the table below. Additionally, 9-14% of poor risk pts remained on therapy for >12 mos and had higher OS (29.8-35.2 mos) and ORR (42-53%). Conclusions: In this large database of poor risk mRCC pts, we highlight that the models have similar predictive accuracy in the targeted therapy era. We also demonstrate that poor risk pts continue to have dismal outcomes and should be considered for trials of novel agents and combinations. A subset of these pts appears to derive prolonged clinical benefit and should be further explored.[Table: see text]

Towards spatially continuous mapping of soil organic carbon in croplands using multitemporal Sentinel-2 remote sensing

In order to fit the small cross-center-biased characteristic of the real world video sequences, an improved version of the well-known cross diamond search algorithm is proposed in this paper. The new algorithm uses a small cross-shaped search patterns in the first two steps to speedup the motion estimation of stationary and quasi-stationary blocks. Experimental results show that this new cross-diamond search algorithm could achieve much higher computational reduction as compared with Diamond Search (DS) and Cross Diamond Search (CDS) while similar prediction accuracy is maintained, and it is especially suitable for videoconferencing sequences.

CD-NuSS: A Web Server for the Automated Secondary Structural Characterization of the Nucleic Acids from Circular Dichroism Spectra Using Extreme Gradient Boosting Decision-Tree, Neural Network and Kohonen Algorithms

BackgroundBreed-specific effects are observed when the same allele of a given genetic marker has a different effect depending on its breed origin, which results in different allele substitution effects across breeds. In such a case, single-breed breeding values may not be the most accurate predictors of crossbred performance. Our aim was to estimate the contribution of alleles from each parental breed to the genetic variance of traits that are measured in crossbred offspring, and to compare the prediction accuracies of estimated direct genomic values (DGV) from a traditional genomic selection model (GS) that are trained on purebred or crossbred data, with accuracies of DGV from a model that accounts for breed-specific effects (BS), trained on purebred or crossbred data. The final dataset was composed of 924 Large White, 924 Landrace and 924 two-way cross (F1) genotyped and phenotyped animals. The traits evaluated were litter size (LS) and gestation length (GL) in pigs.ResultsThe genetic correlation between purebred and crossbred performance was higher than 0.88 for both LS and GL. For both traits, the additive genetic variance was larger for alleles inherited from the Large White breed compared to alleles inherited from the Landrace breed (0.74 and 0.56 for LS, and 0.42 and 0.40 for GL, respectively). The highest prediction accuracies of crossbred performance were obtained when training was done on crossbred data. For LS, prediction accuracies were the same for GS and BS DGV (0.23), while for GL, prediction accuracy for BS DGV was similar to the accuracy of GS DGV (0.53 and 0.52, respectively).ConclusionsIn this study, training on crossbred data resulted in higher prediction accuracy than training on purebred data and evidence of breed-specific effects for LS and GL was demonstrated. However, when training was done on crossbred data, both GS and BS models resulted in similar prediction accuracies. In future studies, traits with a lower genetic correlation between purebred and crossbred performance should be included to further assess the value of the BS model in genomic predictions.

Decreased plasma miR-140-3p is associated with coronary artery disease

A scalable Bayesian functional GWAS method accounting for multivariate quantitative functional annotations with applications for studying Alzheimer disease.