High Performance Computing Center
High Performance Computing Center
News Article | March 14, 2016
The research revealed black phosphorous and monochalcogenide monolayers act differently than any other known 2-D materials at any given temperature because there are four ways to create their atomistic arrangement, and these four arrangements compete and lead to disorder, said Salvador Barraza-Lopez, an assistant professor of physics at the University of Arkansas. "Remarkably, nobody had found that some of these two-dimensional materials become disordered at a room temperature and well before they melt," Barraza-Lopez said. "At the transition temperature the unit cell transforms from a rectangle onto a square and all material properties change." An international research team led by Barraza-Lopez and Pradeep Kumar, assistant professor of physics at the U of A, published its findings in Nano Letters, a journal of the American Chemical Society. The black phosphorous and monochalcogenide monolayers become disordered at a finite temperature, Barraza-Lopez said. "At that moment, the structure transforms from a rectangle to a square and its behavior also changes," he said. Having access to the Trestles supercomputer at the Arkansas High Performance Computing Center was crucial to the study, Barraza-Lopez said. Barraza-Lopez and Mehrshad Mehboudi ran multiple calculations on Trestles for about three weeks each and without interruption. Mehboudi is a doctoral student in the university's interdisciplinary microelectronics-photonics graduate program. "There is no way we could have achieved these results in the timeframe we did without Trestles," Barraza-Lopez said. Explore further: Studies give growers tools to bring new tropical plant to Indiana More information: Mehrshad Mehboudi et al. Two-Dimensional Disorder in Black Phosphorus and Monochalcogenide Monolayers, Nano Letters (2016). DOI: 10.1021/acs.nanolett.5b04613
News Article | March 21, 2016
An international group of physicists has discovered a phenomenon of large magnitude in an unexpected class of materials that can lead to a variety of devices used in optical systems. That phenomenon — the elasto-optic effect — characterizes the formation of a periodic variance of light refraction when an acoustic wave propagates in optical materials, says Yurong Yang, a research assistant professor at the University of Arkansas who led the research. “We found a significantly large elasto-optic effect in thin films made of materials called ferroelectrics,” Yang says, “which are usually considered for their changes in mechanical energy into electrical energy and vice versa, as well in multiferroelectric thin films, which are commonly investigated because of the possible control of their magnetic response by electric input, as well as of their electric response by magnetic input.” The research group published its findings in a paper in Physical Review Letters, the journal of the American Physical Society. A second paper describing the research was published in Nature Communications, an online journal published by the journal Nature. “Those discoveries of a large elasto-optic effect in ferroelectrics and multiferroelectrics therefore broaden the potential of these materials since they can now be put in use to also control their optical responses by elastic property,” says Laurent Bellaiche, Distinguished Professor of physics at the U of A, “which suggests exciting device opportunities arising from this overlooked coupling in these classes of materials.” Yang and Bellaiche, who holds the Twenty-First Century Endowed Professorship in Nanotechnology and Science Education, both conduct research in the Institute for Nanoscience and Engineering and physics department at the U of A. The researchers performed calculations on supercomputers at the Arkansas High Performance Computing Center and a U.S. Department of Defense supercomputing resource. The results published in Physical Review Letters were obtained through a collaborative effort with Zhigang Gui, a U of A physics graduate who is now a postdoctoral research associate at the University of Delaware; Lan Chen and X.K. Meng at Nanjing University in China, and Daniel Sando and Manuel Bibes at University of Paris-Sud in France. The results published in Nature Communications were obtained through a collaborative effort with Daniel Sando and Manuel Bibes and Cecile Carretero, Vincent Garcia, Stephane Fusil, and Agnes Barthelemy at the University of Paris-Sud; Eric Bousquet and Philippe Ghosez at the University of Liege in Belgium; and Daniel Dolfi of Thales Research and Technology in France. Source: University of Arkansas
News Article | February 15, 2017
AMHERST, Mass. - With a new cluster of specialized graphics processing units (GPUs) now installed, the University of Massachusetts Amherst is poised to attract the nation's next crop of top Ph.D. students and researchers in such fields as artificial intelligence, computer vision and natural language processing, says associate professor Erik Learned-Miller of the College of Information and Computer Sciences (CICS). Lead researcher Learned-Miller says, "GPUs are critical for modern computer science research because they have such enormous computational power. They can address extreme computational needs, solving problems 10 times faster than conventional processors, in days rather than months. They can run neural network algorithms that are prohibitively slow on lesser machines. Our new network of 400 GPUs is unusually large for an academic cluster." UMass Amherst's new GPU cluster, housed at the Massachusetts Green High Performance Computing Center in Holyoke, is the result of a five-year, $5 million grant to the campus from Gov. Charlie Baker's administration and the Massachusetts Technology Collaborative last year. It represents a one-third match to a $15 million gift supporting data science and cybersecurity research from the MassMutual Foundation of Springfield. Deep learning research uses neural network algorithms to make sense of large data sets. The approach teaches computers through trial and error to categorize data, much as human brains do. "Deep learning is a revolutionary approach to some of the hardest problems in machine reasoning, and is the 'magic under the hood' of many commercial products and services," says Learned-Miller. "Google Translate, for example, produced more accurate and natural translations thanks to a novel deep-learning approach." Andrew McCallum, professor and founder of the Center for Data Science at UMass Amherst, says, "This is a transformational expansion of opportunity and represents a whole new era for the center and our college. Access to multi-GPU clusters of this scale and speed strengthens our position as a destination for deep learning research and sets us apart among universities nationally." He says the campus currently has research projects that apply deep learning techniques to computational ecology, face recognition, graphics, natural language processing and many other areas. The state funds must be used for computing hardware at UMass Amherst, its Springfield Center for Cybersecurity and for terminals at Mount Holyoke College and the UMass Center in Boston, the researchers note. Learned-Miller says he and colleagues are now in the first year of the grant, during which about $2 million has been spent on two clusters: the GPU cluster dubbed "Gypsum" and a smaller cluster of traditional CPU machines dubbed "Swarm II." Gypsum consists of 400 GPUs installed on 100 computer nodes, along with a storage system and a backup system. It is configured with a leading software package for deploying, monitoring and managing such clusters. Not only do the researchers hope the GPUs will accelerate deep learning research and train a new generation of experts at CICS, but an important overall goal is to foster collaborations between UMass Amherst and industry. For example, if MassMutual data scientists design a practical problem with high computational needs, they can collaborate with sponsored UMass faculty and graduate students to solve it on the Gypsum cluster.
Bertoncini M.,Engineering Ingegneria Informatica |
Pernici B.,Polytechnic of Milan |
Salomie I.,Technical University of Cluj Napoca |
Wesner S.,High Performance Computing Center
Lecture Notes in Business Information Processing | Year: 2011
The vision of the recently started GAMES European Research project is a new generation of energy efficient IT Service Centres, designed taking into account both the characteristics of the applications running in the centre and context-aware adaptivity features that can be enabled both at the application level and within the IT and utility infrastructure. Adaptivity at the application level is based on the service-oriented paradigm, which allows a dynamic composition and recomposition of services to guarantee Quality of Service levels that have been established with the users. At the infrastructure level, adaptivity is being sought with the capacity of switching on and off dynamically the systems components, based on the state of the service centre. However, these two perspectives are usually considered separately, managing at different levels applications and infrastructure. In addition, while performance and cost are usually the main parameters being considered both during design and at run time, energy efficiency of the service centre is normally not an issue. However, given that the impact of service centres is becoming more and more important in the global energy consumption, and that energy resources, in particular in peak periods, are more and more constrained, an efficient use of energy in service centres has become an important goal. In the GAMES project, energy efficiency improvement goals are tackled based on exploiting adaptivity, on building a knowledge base for evaluating the impact of the applications on the service centre energy consumption, and exploiting the application characteristics for an improved use of resources. © 2011 Springer-Verlag Berlin Heidelberg.
News Article | October 23, 2015
Anchor loci described in ref. 12 were extended such that each contained approximately 1,350 bp. In some cases neighbouring loci were joined to form a single locus. Also, loci that performed poorly in ref. 12 were removed from the locus set. This process produced 394 loci (referred to as the version 2 vertebrate loci). Genome coordinates corresponding to these regions in the Gallus gallus genome (galGal3, UCSC genome browser) were identified and sequences corresponding to this region were extracted (coordinates are available in the Zenodo archive (http://dx.doi.org/10.5281/zenodo.28343)). In order to improve the capture efficiency for passerines, we also obtained homologous sequences for Taeniopygia guttata. After aligning the Gallus and Taeniopygia sequences using MAFFT31, alignments were trimmed to produce the final probe region alignments (alignments available in the Zenodo archive), and probes were tiled at approximately 1.5 × tiling density (probe specification will be made available upon publication). Data were collected following the general methods of ref. 12 through the Center for Anchored Phylogenomics at Florida State University (http://www.anchoredphylogeny.com). Briefly, each genomic DNA sample was sonicated to a fragment size of ~150–350 bp using a Covaris E220 focused-ultrasonicator with Covaris microTUBES. Subsequently, library preparation and indexing were performed on a Beckman-Coulter Biomek FXp liquid-handling robot following a protocol modified from ref. 32. One important modification is a size-selection step after blunt-end repair using SPRIselect beads (Beckman-Coulter; 0.9 × ratio of bead to sample volume). Indexed samples were then pooled at equal quantities (typically 12–16 samples per pool), and enrichments were performed on each multi-sample pool using an Agilent Custom SureSelect kit (Agilent Technologies), designed as specified above. After enrichment, the 12 enrichment pools were pooled in groups of three in equal quantities for sequencing on four PE150 Illumina HiSeq2000 lanes (three enrichment pools per lane). Sequencing was performed in the Translational Science Laboratory in the College of Medicine at Florida State University. Paired-read merging (Merge.java). Typically, between 50% and 75% of sequenced library fragments had an insert size between 150 bp and 300 bp. As 150 bp paired-end sequencing was performed, this means that the majority of the paired reads overlap and thus should be merged before assembly. The overlapping reads were identified and merged following the methods of ref. 33. In short, for each degree of overlap for each read we computed the probability of obtaining the observed number of matches by chance, and selected degree of overlap that produced the lowest probability, with a P value less than 10−10 required to merge reads. When reads are merged, mismatches are reconciled using base-specific quality scores, which were combined to form the new quality scores for the merged read (see ref. 33 for details). Reads failing to meet the probability criterion were kept separate but still used in the assembly. The merging process produces three files: one containing merged reads and two containing the unmerged reads. Assembly (Assembler.java). The reads were assembled into contigs using an assembler that makes use of both a divergent reference assembly approach to map reads to the probe regions and a de novo assembly approach to extend the assembly into the flanks. The reference assembler uses a library of spaced 20-mers derived from the conserved sites of the alignments used during probe design. A preliminary match was called if at least 17 of 20 matches exist between a spaced kmer and the corresponding positions in a read. Reads obtaining a preliminary match were then compared to an appropriate reference sequence used for probe design to determine the maximum number of matches out of 100 consecutive bases (all possible gap-free alignments between the read and the reference ware considered). The read was considered mapped to the given locus if at least 55 matches were found. Once a read is mapped, an approximate alignment position was estimated using the position of the spaced 20-mer, and all 60-mers existing in the read were stored in a hash table used by the de novo assembler. The de novo assembler identifies exact matches between a read and one of the 60-mers found in the hash table. Simultaneously using the two levels of assembly described above, the three read files were traversed repeatedly until an entire pass through the reads produced no additional mapped reads. For each locus, mapped reads were then clustered into clusters using 60-mer pairs observed in the reads mapped to that locus. In short, a list of all 60-mers found in the mapped reads was compiled, and the 60-mers were clustered if found together in at least two reads. The 60-mer clusters were then used to separate the reads into clusters for contig estimation. Relative alignment positions of reads within each cluster were then refined in order to increase the agreement across the reads. Up to one gap was also inserted per read if needed to improve the alignment. Note that given sufficient coverage and an absence of contamination, each single-copy locus should produce a single assembly cluster. Low coverage (leading to a break in the assembly), contamination, and gene duplication, can all lead to an increased number of assembly clusters. A whole-genome duplication, for example, would increase the number of clusters to two per locus. Consensus bases were called from assembly clusters as follows. For each site an unambiguous base was called if the bases present were identical or if the polymorphism of that site could be explained as sequencing error, assuming a binomial probability model with the probability of error equal to 0.1 and alpha equal to 0.05. If the polymorphism could not be explained as sequencing error, the ambiguous base was called that corresponded to all of the observed bases at that site (for example, ‘R’ was used if ‘A’ and ‘G’ were observed). Called bases were soft-masked (made lowercase) for sites with coverage lower than five. A summary of the assembly results is presented in a spreadsheet in the electronic data archive (http://dx.doi.org/10.5281/zenodo.28343; Prum_AssemblySummary_Summary.xlsx). Contamination filtering (IdentifyGoodSeqsViaReadsMapped.r, GatherALLConSeqsWithOKCoverage.java). In order to filter out possible low-level contaminants, consensus sequences derived from very low coverage assembly clusters (<10 reads) were removed from further analysis. After filtering, consensus sequences were grouped by locus (across individuals) in order to produce sets of homologues. Orthology (GetPairwiseDistanceMeasures.java, plotMDS5.r). Orthology was then determined for each locus as follows. First, a pairwise distance measure was computed for pairs of homologues. To compute the pairwise distance between two sequences, we computed the percent of 20-mers observed in the two sequences that were found in both sequences. Note that the list of 20-mers was constructed from consecutive 20-mers as well as spaced 20-mers (every third base), in order to allow increased levels of sequence divergence. Using the distance matrix, we clustered the sequences using a neighbour-joining algorithm, but allowing at most one sequence per species to be in a given cluster. Clusters containing fewer than 50% of the species were removed from downstream processing. Alignment (MAFFT). Sequences in each orthologous set were aligned using MAFFT v7.023b31 with “–genafpair” and “–maxiterate 1000” flags. Alignment Trimming (TrimAndMaskRawAlignments3). The alignment for each locus was then trimmed/masked using the following procedure. First, each alignment site was identified as ‘good’ if the most common character observed was present in >40% of the sequences. Second, 20 bp regions of each sequence that contained <10 good sites were masked. Third, sites with fewer than 12 unmasked bases were removed from the alignment. Lastly, entire loci were removed if both outgroups or more than 40 taxa were missing. This filter yielded 259 trimmed loci containing fewer than 2.5% missing characters overall. To minimize the overall model complexity while accurately accounting for substitution processes, we performed a partition-model sensitivity analysis with the development version of PartitionFinder v2.0 (ref. 13), sensu14, and compared a complex partition-model (one partition per gene) to a heuristically optimized (relaxed clustering with the RAxML option for accelerated model selection) partition-model using BIC. Based on a candidate pool of potential partitioning strategies that spanned a single partition for the entire data set to a model allowing each locus to represent a unique partition, the latter approach suggested that 75 partitions of our data set represented the best-fitting partitioning scheme, which reduced the number of necessary model parameters by 71%, and hugely decreased computation time. We analysed each individual locus in RAxML v8.0.20 (ref. 18), and then the concatenated alignment, using the two partitioning strategies identified above with both maximum likelihood and Bayesian based approaches in RAxML v8.0.20, and ExaBayes v1.4.2 9 (ref. 34). For each RAxML analysis, we executed 100 rapid bootstrap inferences and thereafter a thorough ML search using a GTR+Γ model of nucleotide substitution for each data set partition. Although this may potentially over-parameterize a partition with respect to substitution model, the influence of this form of model over-parameterization has been found to be negligible in phylogenetic inference35. For the Bayesian analyses, we ran four Metropolis-coupled ExaBayes replicates for 10 million generations, each with three heated chains, and sampling every 1,000 generations (default tuning and branch swap parameters; branch lengths among partitions were linked). Convergence and proper sampling of the posterior distribution of parameter values were assessed by checking that the effective sample sizes of all estimated parameters and branch lengths were greater than 200 in the Tracer v1.6 software36 (most were greater than 1,000), and by using the ‘sdsf’ and ‘postProcParam’ tools included with the ExaBayes package to ensure the average standard deviation of split frequencies and potential scale reduction factors across runs were close to zero and one, respectively. Finally, to check for convergence in topology and clade posterior probabilities, we summarized a greedily refined majority-rule consensus tree (default) from 10,000 post burn-in trees using the ExaBayes ‘consense’ tool for each run independently and then together. Analyses of the reduced data set referenced in the main text were conducted using the same partition-model as the full data set. To explore variation in gene tree topology and to look for outliers that might influence combined analysis, we calculated pairwise Robinson-Foulds37 (RF) and Matching Splits (MS) tree distances implemented in TreeCmp38. We then visualized histograms of tree distances and multidimensional scaling plots in R, and estimated neighbour-joining ‘trees-of-trees’ in the Phangorn R package sensu lato39, 40. Using RF and MS distances, outlier loci were identified as those that occurred in the top 10% of pairwise distances for >30 comparisons to other loci (~10%) in the data set. We also identified putative outlier loci using the kdetrees.complete function of the kdetrees R package41. All three methods identified 13 of the same loci as potential outliers; however removal of these loci from the analysis had no effect on estimating topology or branch lengths. Although fully parametric estimation (for example, *BEAST, see ref. 42) of a coalescent species tree with hundreds of genes and hundreds of taxa is not currently possible, we estimated species trees using three gene-tree summation methods that have been shown to be statistically consistent under the multispecies coalescent model43. First, we used the STRAW web server44 to estimate bootstrapped species trees using the STAR19 and NJ-ST20 algorithms (also available through STRAW). The popular MP-EST45 method cannot currently work for more than ~50 taxa. STAR takes rooted gene trees and uses the average ranks of coalescence times19 to build a distance matrix from which a species tree is computed with the neighbour-joining method46. By contrast, NJst applies the neighbour-joining method to a distance matrix computed from average gene-tree internode distances, and relaxes the requirement for input gene trees to be rooted20. We also summarized a species tree with the ASTRAL 4.7.6 algorithm. With simulated data, ASTRAL has been shown to outperform concatenation or other summary methods under certain amounts of incomplete lineage sorting21. For very large numbers of taxa and genes, ASTRAL uses a heuristic search to find the species tree that agrees with the largest number of quartet trees induced by the set of input gene trees. For analysis with ASTRAL, we also attempted to increase the resolution of individual gene trees (Supplementary Fig. 2) by generating supergene alignments using the weighted statistical binning pipeline of refs 47, 48 with a bootstrap score of 0.75 as a bin threshold. STAR, NJst (not shown), and the binned ASTRAL (Supplementary Fig. 3) analysis produced virtually identical inferences when low support branches (<0.75) were collapsed, and differed only with respect to the resolution of a few branches. NJst resolved the Passeroidea (Fringilla plus Spizella) as the sister group to a paraphyletic sample of Sylvioidea (Calandrella, Pycnonotus, and Sylvia), while STAR does not resolve this branch. Comparing STAR/NJst to ASTRAL, we find five additional differences: (1) within tinamous, STAR/NJst resolves Crypturellus as sister to the rest of the tinamous, whereas ASTRAL resolves Crypturellus as sister to Tinamus (similar to ExaBayes/RAxML); (2) STAR/NJst resolves pigeons as sister to a clade containing Mesitornithiformes and Pteroclidiformes, while ASTRAL does not resolve these relationships; (3), STAR/NJst fails to resolve Oxyruncus and Myiobius as sister genera, while ASTRAL does (similar to RAxML/ExaBayes); (4), in STAR/NJst, bee-eaters (Merops) are resolved as the sister group to coraciiforms (congruent with ref. 4), while ASTRAL resolves bee-eaters as sister to the rollers (Coracias) (similar to RAxML/ExaBayes); (5) lastly, in STAR/NJst, buttonquail (Turnix) is resolved as sister to the most inclusive clade of Charadriiformes not including Burhinus, Charadrius, Haematopus, and Recurvirostra, while in ASTRAL, buttonquail is resolved as sister to a clade containing Glareola, Uria, Rynchops, Sterna, and Chroicocephalus (similar to RAxML/ExaBayes). Although lower level relationships detected with concatenation are generally recapitulated in the species trees, few of the higher level, or interordinal, relationships are resolved. This lack of resolution of the gene-tree species-tree based inferences relative to the inferences based on concatenation are not surprising, as it is increasingly recognized that the phylogenetic information content required to resolve the gene-tree histories of individual loci becomes scant at deep timescales47. Despite our extensive taxon sampling and the slow rate of nucleotide substitution that characterizes loci captured using anchored enrichment12, no single locus was able to fully resolve a topology, and this lack of information will challenge the accuracy of any coalescent-based summary approach relative to concatenation49, 50, 51, 52, 53, 54. Finally, all summation methods tested here assume a priori that the only source of discordance among gene trees is deep coalescence, and violations of this assumption may introduce systematic error in phylogeny estimation54. Site-specific evolutionary rates, λ , were calculated for each locus using the program HyPhy55 in the PhyDesign web interface56 in conjunction with a guide chronogram generated by a nonparametric rate smoothing algorithm57 applied to our concatenated RAxML tree. Using these rates to predict whether an alignment will yield correct, incorrect, or no resolution of a given node, we quantified the probability of phylogenetically informative changes (ψ)16 contributing to the resolution of the earliest divergences in Neoaves. Estimates generated under a three character state model58 reveal that the majority of loci have a strong probability of ψ, and suggest a high potential for most loci and partitions containing multiple loci (assigned by PartitionFinder) to correctly resolve this internode. The potential for resolution as a consequence of phylogenetic signal is therefore high relative to the potential for saturation and misleading inference induced by stochastic changes along the subtending lineages (Supplementary Fig. 4a). To assess the information content of the loci across the entire topology, we profiled their phylogenetic informativeness (PI)15, (Supplementary Fig. 4b). There was considerable variation in PI across loci (Supplementary Fig. 4). In all cases, the loci with the lowest values of ψ are categorized by substantially lower (60–90%) values of PI, rather than sharp declines in their PI profiles. The absence of a sharp decline in the PI profile suggests that a lack of phylogenetic information, rather than rapid increases in homoplasious sites, underlie low values of the probability of signal ψ59. Because declines in PI can be attributed to increases in homoplasious site patterns59, we further assessed the phylogenetic utility of data set partitions by quantifying the ratio of PI at the most recent common ancestor of Neoaves to the PI at the most recent common ancestor of Aves (Supplementary Fig. 4c). Values of this ratio that are less than 1 correspond to a rise in PI towards the root. Values close to 1 correspond to fairly uniform PI. Values greater than 1 correspond to a decline in PI towards the root. Sixty-six out of 75 partitions demonstrated less than a 50% percent decline in PI, and only six partitions demonstrated a decline of PI greater than 75% (Supplementary Fig. 4c). As all but a few nodes in this study represent divergences younger than the crown of Neoaves, these ratios of PI suggest that the predicted impact of homoplasy on our topological inferences should be minimal. As PI profiles do not directly predict the impact of homoplasious site patterns on topological resolution16, 60, we evaluated probabilities of ψ for focal nodes using both the concatenated data set as well as individual loci that span the variance in locus lengths. Concordant with expectations from the PI profiles, all quantifications strongly support the prediction that homoplasy will have a minimal impact on topological resolution for the concatenated data set across a range of tree depths and internode distances (ψ = 1.0 for all nodes), while individual loci vary in their predicted utility (Supplementary Fig. 4d). As the guide tree does not represent a true known tree, we additionally quantified ψ across a range of tree depths and internode distances to test if our predictions of utility are in line with general trends in the data. Concordant with our results above, the concatenated data set is predicted to be of high phylogenetic utility at all timescales (ψ = 1.0 for all nodes), while the utility of individual loci begins to decline for small internodes at deep tree depths (Supplementary Fig. 5). We estimated a time-calibrated tree with a node dating approach in BEAST 1.8.1 (ref. 42) that used 19 well justified fossil calibrations phylogenetically placed by rigorous, apomorphy-based diagnoses (see the descriptions of avian calibration fossils in the Supplementary Information). We used a starting tree topology based on the ExaBayes inference (Fig. 1), and prior node age calibrations that followed a lognormal parametric distribution based on occurrences of fossil taxa. To prevent BEAST from exploring topology space and only allow estimation of branch lengths, we turned off the subtree-slide, Wilson–Balding, and narrow and wide exchange operators61, 62. Finally, we applied a birth–death speciation model with default priors. As rates of molecular evolution are significantly variable across certain bird lineages63, 64, 65, we applied an uncorrelated relaxed clock (UCLN) to each partition of the data set where rates among branches are distributed according to a lognormal distribution66. All dating analyses were performed without crocodilian outgroups to reduce the potential of extreme substitution rate heterogeneity to bias rate and consequent divergence time estimates of the UCLN model67. All calibrations were modelled using soft maximum age bounds to allow for the potential of our data to overwhelm our user-specified priors68. Soft maximum bounds are the preferred method for assigning upper limits on the age of phylogenetic divergences69. As effective priors necessarily reflect interactions between user specified priors, topology, and the branching-model, they may not precisely reflect the user-specified priors70. To correct for this potential source of error, we carefully examined the effective calibration priors by first running the prepared BEAST XML without any nucleotide data (until all ESS values were above 200). We then iteratively adjusted our user-defined priors until all of the effective priors (as examined in the Tracer software) reflected the intended calibration densities. Finally, using the compare.phylo function in the Phyloch R package, we examined how the inclusion of molecular data influenced the divergence time estimates relative to the effective prior (Supplementary Fig. 9; see below). Our initial approach was to set a prior’s offset to the age of its associated fossil; the mean was then manually adjusted such that 95% of the calibration density fell more recently than the K–Pg boundary at 65 Ma (million years ago) (the standard deviation was fixed at 1 Ma). In general, priors constructed this way generated calibration densities that specified their highest density peak (their mode) about 3–5 million years older than the age of the offset. We applied a loose gamma prior to the node reflecting the most recent common ancestor of crown birds—we used an offset of 60.5 Ma (the age of the oldest known definitive, uncontroversial crown bird fossil; the stem penguin Waimanu), and adjusted the scale and shape of the prior such that 97.5% of the calibration density fell more recently than 86.5 Ma71 (see below and Supplementary Information for discussion of the >65 Ma putative crown avian Vegavis). This date (86.5 Ma) reflects the upper bound age estimate of the Niobrara Formation—one of many richly fossiliferous Mesozoic deposits exhibiting many crownward Mesozoic stem birds, without any trace of avian crown group representatives. The Niobrara, in particular, has produced hundreds of stem birds and other fragile skeletons, without yielding a single crown bird fossil, and therefore represents a robust choice for a soft upper bound for the root divergence of the avian crown71, 72, 73. Previous soft maxima employed for this divergence have arbitrarily selected the age of other Mesozoic stem avians (that is, Gansus yumenensis, 110 Ma) that are phylogenetically stemward of the Niobrara taxa28. Although the implementation of very ancient soft maxima such as the age of Gansus are often done in the name of conservatism, the extremely ancient divergence dates yielded by such analyses illustrate the misleading influence of assigning soft maxima that are vastly too old to be of relevance to the divergence of crown group birds74. However, this problem has been eliminated in some more recent analyses75. All of the fossil calibrations employed in our analysis represent neognaths; rootward divergences within Aves (for example the divergence between Palaeognathae and Neognathae, and Galloanserae and Neoaves) cannot be confidently calibrated due to a present lack of fossils representing the palaeognath, neognath, galloanserine, and neoavian stem groups. As such, the K–Pg soft bound was only applied to comparatively apical divergences within neognaths. Although the question of whether major neognath divergences occurred during the Mesozoic has been the source of controversy76, 77, 78, renewed surveys of Mesozoic sediments for definitive crown avians or even possible crown neoavians have been unsuccessful (with the possible exception of Vegavis; see Supplementary Information), and together with recent divergence dating analyses have cast doubt on the presence of neoavian subclades before the K–Pg mass extinction1, 74, 79. Further, recent work has demonstrated the tendency of avian divergence estimates to greatly exceed uninformative priors, resulting in spuriously ancient divergence dating results (for example, refs 28, 75, 76, 80). These results motivated our implementation of the 65 Ma soft bound for our neoavian calibrations. Contrary to expectation, when we compared the effective prior on the entire tree to the final summary derived from the posterior distribution of divergence times (Supplementary Fig. 9), we found no overall trend of posterior estimated ages post-dating prior calibrations. In fact, the inclusion of our molecular data decreases the inferred ages of almost all of the deepest nodes in our tree. A similar result has been obtained for mammals by using large amounts of nuclear DNA sequences81. Future work investigating the interplay of the density of genomic sampling and the application of various calibration age priors will be indispensible for sensitivity analyses to help us further develop a robust timescale of avian evolution. However, the pattern of posterior versus prior age estimates observed in our study raises the prospect that the new class of data used in this study (that is, semi-conserved anchor regions) may exhibit some immunity to longstanding problems associated with inferring avian divergence times, such as systematically over-estimating the antiquity of extant avian clades. In addition to making predictions about the phylogenetic utility of a locus or partition towards topological resolution, PI profiles have recently also been used to mitigate the influence of substitution saturation on divergence time estimates82. Given the variance in PI profile shapes for captured loci and their subsequent partition assignments (Supplementary Fig. 4c), and observations that alignments and subsets of data alignments characterized by high levels of homoplasy can mislead branch length estimation83, 84, we limited our divergence time estimates to 36 partitions that did not exhibit a decline in informativeness towards the root of the tree. We ran BEAST on each partition separately until parameter ESS values were greater than 200 (most were greater than 1,000) to ensure adequate posterior sampling of each parameter value. After concatenating 10,000 randomly sampled post burn-in trees from each of these completed analyses, we summarized a final MCC tree with median node heights in TreeAnnotator v1.8.1 (ref. 42). Supplementary Fig. 6 shows the full, calibrated Bayesian tree (Fig. 1) with 95% HPD confidence intervals on the node ages, and Supplementary Fig. 7 shows the distribution of estimated branching times, ranked by median age (using clade numbers from Fig. 1). All computations were carried out on 64-core PowerEdge M915 nodes on the Louise Linux cluster at the Yale University Biomedical High Performance Computing Center. No statistical methods were used to predetermine sample size.
News Article | November 15, 2016
PITTSBURGH, Nov. 15, 2016 /PRNewswire/ -- ANSYS (NASDAQ: ANSS), the High Performance Computing Center (HLRS) of the University of Stuttgart and Cray Inc. have set a new supercomputing world record by scaling ANSYS® Fluent® to over 172,000 computer cores on the HLRS supercomputer Hazel...
News Article | November 15, 2016
SEATTLE and SALT LAKE CITY, Nov. 15, 2016 (GLOBE NEWSWIRE) -- Global supercomputer leader Cray Inc. (Nasdaq:CRAY) today announced new customers for the Cray® Urika-GX™ system at the 2016 Supercomputing Conference in Salt Lake City, Utah, and highlighted expanded capabilities for the agile analytics platform coming in December. Cray continues to build momentum for the Cray Urika-GX system, which fuses supercomputing technologies with an open, enterprise-ready software framework for big data analytics. Cray was recently awarded contracts for Cray Urika-GX system from customers specializing in manufacturing, and customer engagement. These new customers include the High Performance Computing Center of the University of Stuttgart (HLRS), and Phizzle – a customer engagement marketing and Internet of Things (IoT) company that enables the ingestion, analysis and reaction to big data in real time. The new contracts follow a previously announced collaboration with Deloitte Advisory Cyber Risk Services to offer threat analytics services powered by the Cray Urika-GX system. “Cray has a long history of partnering with customers as they face complex and evolving analytics challenges,” said Ryan Waite, Cray’s senior vice president of products. “The steady adoption of Urika-GX from a broader set of customers validates the benefits from integrating supercomputing hardware with powerful, open source analytics software like Apache Spark, to face those challenges. Our December update will provide additional support for enterprise customers with broader storage connectivity and broader data visualization support.” At HLRS, a project is underway to combine high performance computing and high performance data analytics to expand the Center’s capabilities in automotive manufacturing. HLRS, which works in close collaboration with its industrial partners in the aerospace and automotive industries, will acquire two Urika-GX systems to help drive more predictive maintenance by implementing IoT style analytics with complex, large-scale modeling workloads. “In the automotive industry, high performance computing and data analytics play an important role in product development, and with the progress of digitization, even larger and more complex data sets are generated that cannot be analyzed using conventional methods,” said Prof. Dr. Michael Resch, director of HLRS. “In cooperation with Cray and industrial users in the region, we are testing the possible applications of hardware in industrial environments, and the Urika-GX system will play an important role for us as we work to find a practical solution.” The collaboration between Cray and Phizzle will allow software developers on its phz.io platform to support customers at greater scale and speeds by achieving a 16x performance increase over their existing infrastructure. Developers will be able to leverage Phizzle’s phz.io platform on its Cray Urika-GX system to gain supercomputing speeds that have previously been too complex or expensive to access. “Powered by the Cray Urika-GX, phz.io will change the way everyday developers interact with their consumers,” said Stephen Goldberg, CTO of Phizzle. “By putting supercomputing and analytics in the hands of mobile, enterprise, and application developers who traditionally would not have access to Cray, phz.io allows developers to interact and react at speeds never seen before.” Cray is also introducing new software features to further enable enterprise customers. With the upcoming software release in December, Cray Urika-GX customers will be able to take advantage of a broader range of enterprise storage, including GPFS (General Parallel File System) and NFS (Network File System). Additionally, the Cray Urika-GX system will enable customers to leverage Tableau visualization tools and the latest version of Spark 2.0. The Cray Urika-GX system features Intel® Xeon® E5 v4 processors, up to 22 terabytes of DRAM memory, up to 176 terabytes of local SSD storage capacity, and the high speed Aries network interconnect, which together provide the unmatched in-memory compute and network performance necessary to solve the most demanding big data problems. An exclusive feature of the Cray Urika-GX system is the Cray Graph Engine for fast, complex pattern matching and iterative discovery. With the Cray Graph Engine, customers can tackle multi-terabyte datasets comprised of billions of objects to uncover hidden relationships in even the nosiest of data. The Cray Graph Engine can run in conjunction with open analytics tools such as Hadoop and Spark, enabling customers to build complete end-to-end analytics workflows and avoid unnecessary data movement. For more information on the Cray Urika-GX system, please visit the Cray website at www.cray.com. About Cray Inc. Global supercomputing leader Cray Inc. (Nasdaq:CRAY) provides innovative systems and solutions enabling scientists and engineers in industry, academia and government to meet existing and future simulation and analytics challenges. Leveraging more than 40 years of experience in developing and servicing the world’s most advanced supercomputers, Cray offers a comprehensive portfolio of supercomputers and big data storage and analytics solutions delivering unrivaled performance, efficiency and scalability. Cray’s Adaptive Supercomputing vision is focused on delivering innovative next-generation products that integrate diverse processing technologies into a unified architecture, allowing customers to meet the market’s continued demand for realized performance. Go to www.cray.com for more information. Safe Harbor Statement This press release contains forward-looking statements within the meaning of Section 21E of the Securities Exchange Act of 1934 and Section 27A of the Securities Act of 1933, including, but not limited to, statements related to its product development plans, including the timing and availability of updates and new features for Cray Urika-GX systems, the sales prospects of Cray Urika-GX systems, the ability of Cray Urika-GX systems to perform as expected and Cray’s ability to deliver systems that meets the requirements of HLRS and Phizzle. These statements involve current expectations, forecasts of future events and other statements that are not historical facts. Inaccurate assumptions and known and unknown risks and uncertainties can affect the accuracy of forward-looking statements and cause actual results to differ materially from those anticipated by these forward-looking statements. Factors that could affect actual future events or results include, but are not limited to, the risk that Cray is not able to successfully complete its planned product development efforts related to the Cray Urika-GX within the planned timeframes or at all, the risk that Cray Urika-GX systems do not perform as expected or as required by customers or partners, the risk that Cray will not be able to sell Cray Urika-GX systems as expected, the risk that the systems required by HLRS and Phizzle are not delivered in a timely fashion or do not perform as expected and such other risks as identified in the Company’s quarterly report on Form 10-Q for the quarter ended September 30, 2016, and from time to time in other reports filed by Cray with the U.S. Securities and Exchange Commission. You should not rely unduly on these forward-looking statements, which apply only as of the date of this release. Cray undertakes no duty to publicly announce or report revisions to these statements as new information becomes available that may change the Company’s expectations. Cray, and the stylized CRAY mark are registered trademarks of Cray Inc. in the United States and other countries, and Urika-GX is a trademark of Cray Inc. Other product and service names mentioned herein are the trademarks of their respective owners.
News Article | December 23, 2016
SUNNYVALE, Californie--(BUSINESS WIRE)--Le HPC Advisory Council (HPCAC), une organisation de premier plan travaillant pour la recherche, l'externalisation et l'éducation en matière de calcul haute performance (high-performance computing, HPC, ou CHP en français) qui a pour vocation de promouvoir la collaboration internationale ainsi que la recherche et les technologies révolutionnaires, a annoncé aujourd'hui les dates clés de sa série de conférences internationales qui auront lieu en 2017 aux États-Unis et en Suisse. Ces conférences ont pour but d'attirer une participation à l'échelle de la communauté, des commanditaires leaders du secteur et des experts du CHP. La série 2017 commencera avec la 7e Conférence annuelle de Stanford, qui se tiendra les 7 et 8 février 2017 à Stanford, en Californie, dans le Hall Paul Brest du Munger Conference Center au sein du campus principal de l'université. La conférence est gratuite et les participants doivent s'inscrire avant le 1er février 2017. La date limite des propositions de soumission est le 31 décembre 2016 et celles-ci peuvent être soumises sur les pages dédiées aux conférences du site Web du HPCAC. La 8e Conférence suisse annuelle se tiendra du 10 au 12 avril 2017 à Lugano, en Suisse, au Palazzo dei Congressi. Les frais d'inscription de la conférence sont de 130 CHF et couvrent les pauses-café et les déjeuners durant la conférence de trois jours, ainsi qu'une excursion spéciale en groupe. L'inscription est ouverte jusqu'au 31 mars 2017. Les participants peuvent bénéficier d'un tarif préférentiel de 80 CHF en s'inscrivant avant le 31 janvier 2017. La date limite des propositions de soumission est le 2 février 2017 et celles-ci peuvent également être soumises sur les pages dédiées aux conférences du site Web du HPCAC. Chaque année, des organisations internationales, des institutions de recherche de premier plan ainsi que des entreprises du secteur privé basées aux États-Unis, en Europe et en Chine s'associent et collaborent avec le HPCAC pour co-héberger et participer à ces forums immersifs dont l'objectif est de présenter des praticiens experts et des solutions révolutionnaires rendues possibles par les disciplines et les innovations technologiques dans le domaine du CHP. Chaque conférence est organisée grâce à une collaboration directe avec des contributeurs qui représentent l'ensemble de la communauté mondiale de la recherche et du développement, et qui contribuent tous généreusement de leur temps et de leur expertise. Les conférences proposent de vastes ordres du jour qui sont développés conjointement par des experts du secteur et via des appels ouverts à des soumissions de communications et de thèmes. Les contributeurs sont libres de définir le format de présentation de leur choix et les sessions peuvent être soumises en tant que présentations individuelles ou en tant que séries ; session(s) technique(s), tutoriel(s), atelier(s), etc. Les contributeurs sont également libres de classer leurs soumissions d'une série en tant que sessions introductives, intermédiaires ou avancées. « Le CHP évolue constamment et reflète la force motrice qui sous-tend de nombreuses percées médicales, industrielles et scientifiques basées sur la recherche et qui exploitent la puissance du CHP. Pourtant, nous n'avons fait que gratter la surface en ce qui concerne l'exploitation des opportunités illimitées que présentent le CHP, la modélisation et la simulation », a déclaré Gilad Shainer, président du HPC Advisory Council. « La série de conférences du HPCAC est une occasion unique pour la communauté mondiale du CHP de se rassembler comme jamais auparavant afin partager, de collaborer et d'innover pour l'avenir. » « L'enthousiasme et le support dont nous bénéficions chaque année en réunissant des personnes du secteur universitaire, de la communauté de la technologie et du monde entier en tant que présentateurs, participants et homologues a pour effet de changer radicalement nos vies et nos carrières », a ajouté Steve Jones, directeur du High Performance Computing Center de Stanford. « La conférence encourage le partage collaboratif des percées et des meilleures pratiques de recherche à un niveau sans précédent pour en faire bénéficier, non seulement les participants, mais également à terme l'humanité. En éduquant et en partageant dans toute la communauté du CHP, nous finissons par en tirer des leçons et de nouvelles connaissances à l'heure où nous nous efforçons tous de créer un monde meilleur. » « Qu'il s'agisse de comprendre les origines de l'univers ou de cartographier le génome humain, les êtres humains ont une soif insatiable de comprendre et de repousser les limites de la technologie et de l'innovation », a confié pour sa part Hussein El Harake, directeur du HPCAC Switzerland Center of Excellence et ingénieur système CSCS. « Cette conférence nous à aidé à élargir nos collaborations et à développer des liens plus solides entre la science, l'industrie et le monde universitaire, afin de créer un écosystème CHP plus uni et d'aller continuellement de l'avant dans notre compréhension collective de ce que le CHP peut accomplir et de ce qu'il accomplit dans les faits. » Les partenaires du secteur, membres du HPCAC et experts techniques travaillent tous de concert pour rapprocher le secteur privé et les institutions publiques dans le but de faciliter ces forums internationaux. Bien que les conférences soient des évènements à but non lucratif, le rendement global de l'investissement est énorme. Les hôtes, commanditaires et participants aux conférences sont tous gagnants parce qu'ils sont informés des tous derniers développements, et qu'ils peuvent améliorer leurs connaissances et leurs perspectives, ainsi qu'accéder plus tôt à des technologies nouvelles et émergentes, des services, des capacités, des résultats partagés, des meilleures pratiques, et bien plus. Fondé en 2008, le HPC Advisory Council est une organisation internationale à but non lucratif de plus de 400 membres engagés pour l'éducation et la sensibilisation. Ses membres partagent leur expertise, dirigent des groupes d'intérêt spéciaux et ont accès au centre technologique HPCAC pour explorer les opportunités et promouvoir les avantages des technologies, des applications et du développement futur du calcul haute performance. Le HPC Advisory Council organise de multiples conférences et concours STEM internationaux, notamment le Concours étudiant RDMA en Chine et le Concours de groupements d'étudiants lors des conférences annuelles ISC High Performance. L'adhésion est gratuite et facultative. Pour plus d'informations, voir: www.hpcadvisorycouncil.com
News Article | December 22, 2016
SUNNYVALE, Calif.--(BUSINESS WIRE)--The HPC Advisory Council (HPCAC), a leading organization for high-performance computing (HPC) research, outreach and education, an organization dedicated to furthering global collaborations, breakthrough research and technologies, today announced key dates for its 2017 international conference series in the USA and Switzerland. The conferences are designed to attract community-wide participation, industry leading sponsors and subject matter experts. The 2017 series begins with the 7th annual Stanford Conference on Feb. 7-8 2017, Stanford, CA, at Munger Conference Center’s Paul Brest Hall on the university’s main campus. The conference is free of charge and attendees must register by Feb. 01, 2017. Deadline for submission proposals is Dec. 31, 2016 and can be submitted via the conference pages on the HPCAC website. The 8th annual Swiss Conference will be held April 10-12 2017, Lugano, Switzerland, at the Palazzo dei Congressi. The conference charges a nominal fee of 130 CHF and includes breaks and lunch during the three day conference along with a special group outing. Registration is open through Mar. 31, 2017. Participants can take advantage of early bird pricing of 80 CHF by registering before Jan. 31, 2017. Deadline for submission proposals is Feb. 2, 2017 and can also be submitted via the conference pages on the HPCAC website. The 2017 schedule will also include ongoing events in China and Spain, with those dates to be announced separately. Sponsorship information for the conference series is available via the HPCAC website. Every year, leading international organizations, research institutions and private industry in the U.S., Europe and China partner and collaborate with the HPCAC to co-host and participate in these immersive forums that strive to showcase expert practitioners and breakthrough solutions ̶ all made possible by HPC disciplines and technology innovations. Each conference is organized through direct collaboration with contributors from across the global research and development community, all of whom generously contribute their time and expertise. The conferences offer expansive agendas that are collaboratively developed by industry experts as well as through the open calls for submissions of papers and topics. Contributors have the freedom to define their preferred delivery format and sessions can be submitted as lone presentations or a series; technical session(s), tutorial(s), workshop(s), etc. Contributors are also free to classify their submissions of a series as introductory, intermediate or advanced sessions. “HPC is constantly evolving and reflects the driving force behind many medical, industrial and scientific breakthroughs using research that harnesses the power of HPC and yet, we’ve only scratched the surface with respect to exploiting the endless opportunities that HPC, modeling, and simulation present,” said Gilad Shainer, chairman of the HPC Advisory Council. “The HPCAC conference series presents a unique opportunity for the global HPC community to come together in an unprecedented fashion to share, collaborate, and innovate our way into the future.” “The enthusiasm and support we experience each year as we bring people together from across campus, the tech community and the globe as presenters, participants and peers is nothing short of life and career altering,” said Steve Jones, Director of Stanford’s High Performance Computing Center. “The conference encourages collaborative sharing of research breakthroughs and best practices on an unprecedented level benefitting not just attendees but ultimately, humanity. In educating and sharing across the HPC community, we ultimately take away lessons learned and newfound knowledge as we all strive to make the world a better place.” “From understanding the beginnings of the universe to mapping the human genome, human beings possess an insatiable drive to understand and stretch the boundaries of technology and innovation,” said Hussein El Harake, Director of the HPCAC Switzerland Center of Excellence and CSCS System Engineer. “This conference has helped expand our collaborations and develop stronger bonds between science, industry and academia, creating a tighter HPC ecosystem and continuously moving forward in our collective understanding regarding what HPC can and is accomplishing.” Industry partners, HPCAC members and technical experts all work closely together to bring private industry together with public institutions to facilitate these international forums. While the conferences are non-profit events, the overall return on investment is formidable. Conference hosts, sponsors and attendees all benefit from learning about the latest developments, increased knowledge and insights as well as early access to new and emerging technologies, services, capabilities, shared results, best practices and more. Founded in 2008, the non-profit HPC Advisory Council is an international organization with over 400 members committed to education and outreach. Members share expertise, lead special interest groups and have access to the HPCAC technology center to explore opportunities and evangelize the benefits of high performance computing technologies, applications and future development. The HPC Advisory Council hosts multiple international conferences and STEM challenges including the RDMA Student Competition in China and the Student Cluster Competition at the annual ISC High Performance conferences. Membership is free of charge and obligation. More information: www.hpcadvisorycouncil.com
Patel N.,High Performance Computing Center |
Schneider R.,High Performance Computing Center |
Kuster U.,High Performance Computing Center
Proceedings of the IEEE Southwest Symposium on Image Analysis and Interpretation | Year: 2016
The most significant constituent for a realistic blood flow simulation is the precise extraction of the patient specific geometry from digital computer tomographic images. The reconstructed data sets are not always sufficient to extract the contours/surfaces of the large aortic structures to build the computational model directly. The raw images are processed with the help of low level filters to remove in-homogeneity and undesired noise to achieve a smoother contour of the desired structures close to the original image data-set. A spatial filter, based on patch size, is applied iteratively on an image. It will be shown that with large patch sizes, distortion increases and with smaller patch sizes, a smooth contour is not achieved. A new, adaptive approach based on the image gradient which provides intrinsic information of the underlying objects in an image is presented. This new approach helps to apply the filters locally in a controlled manner to achieve the desired smooth contour with minimum distortion. © 2016 IEEE.