Google Compute Engine offers a high-performance, costeffective means for running I/O-intensive applications. This report details our experience running large-scale, highperformance sorting jobs on GCE. We run sort applications up to 100 TB in size on clusters of up to 299 VMs, and find that we are able to sort data at or near the hardware capabilities of the locally attached SSDs. In particular, we sort 100 TB on 296 VMs in 915 seconds at a cost of $154.78. We compare this result to our previous sorting experience on Amazon Elastic Compute Cloud and find that Google Compute Engine can deliver similar levels of performance. Although individual EC2 VMs have higher levels of performance than GCE VMs, permitting significantly smaller cluster sizes on EC2, we find that the total dollar cost that the user pays on GCE is 48% less than the cost of running on EC2.

Cloud computing providers have recently begun to offer high-performance virtualized flash storage and virtualized network I/O capabilities, which have the potential to increase application performance. Since users pay for only the resources they use, these new resources have the potential to lower overall cost. Yet achieving low cost requires choosing the right mixture of resources, which is only possible if their performance and scaling behavior is known. In this paper, we present a systematic measurement of recently introduced virtualized storage and network I/O within Amazon Web Services (AWS). Our experience shows that there are scaling limitations in clusters relying on these new features. As a result, provisioning for a large-scale cluster differs substantially from small-scale deployments. We describe the implications of this observation for achieving efficiency in large-scale cloud deployments. To confirm the value of our methodology, we deploy cost-efficient, high-performance sorting of 100 TB as a large-scale evaluation.

We present TritonSort, a sorting system designed to maximize system resource utilization. We present the results for: Indy GraySort, Daytona GraySort, Indy MinuteSort, Indy CloudSort, and Daytona CloudSort.

We present TritonSort, a highly efficient, scalable sorting system. It is designed to process large datasets, and has been evaluated against as much as 100TB of input data spread across 832 disks in 52 nodes at a rate of 0.938TB/min. When evaluated against the annual Indy GraySort sorting benchmark, TritonSort is 66% better in absolute performance and has over six times the per-node throughput of the previous record holder. When evaluated against the 100TB Indy JouleSort benchmark, TritonSort sorted 9703 records/Joule. In this article, we describe the hardware and software architecture necessary to operate TritonSort at this level of efficiency. Through careful management of system resources to ensure cross-resource balance, we are able to sort data at approximately 80% of the disks’ aggregate sequential write speed. We believe the work holds a number of lessons for balanced system design and for scale-out architectures in general. While many interesting systems are able to scale linearly with additional servers, per-server performance can lag behind per-server capacity by more than an order of magnitude. Bridging the gap between high scalability and high performance would enable either significantly less expensive systems that are able to do the same work or provide the ability to address significantly larger problem sets with the same infrastructure.

"Big Data" computing increasingly utilizes the MapReduce programming model for scalable processing of large data collections. Many MapReduce jobs are I/O-bound, and so minimizing the number of I/O operations is critical to improving their performance. In this work, we present Themis, a MapReduce implementation that reads and writes data records to disk exactly twice, which is the minimum amount possible for data sets that cannot fit in memory. In order to minimize I/O, Themis makes fundamentally different design decisions from previous MapReduce implementations. Themis performs a wide variety of MapReduce jobs -- including click log analysis, DNA read sequence alignment, and PageRank -- at nearly the speed of TritonSort's record-setting sort performance [29].

We present TritonSort, a highly efficient, scalable sorting system. It is designed to process large datasets, and has been evaluated against as much as 100 TB of input data spread across 832 disks in 52 nodes at a rate of 0.916 TB/min. When evaluated against the annual Indy GraySort sorting benchmark, TritonSort is 60% better in absolute performance and has over six times the per-node efficiency of the previous record holder. In this paper, we describe the hardware and software architecture necessary to operate TritonSort at this level of efficiency. Through careful management of system resources to ensure cross-resource balance, we are able to sort data at approximately 80% of the disks' aggregate sequential write speed. We believe the work holds a number of lessons for balanced system design and for scale-out architectures in general. While many interesting systems are able to scale linearly with additional servers, per-server performance can lag behind per-server capacity by more than an order of magnitude. Bridging the gap between high scalability and high performance would enable either significantly cheaper systems that are able to do the same work or provide the ability to address significantly larger problem sets with the same infrastructure.

The Self-Appraisal Questionnaire (SAQ) is a 72-item self-report measure designed to predict violent and nonviolent recidivism among adult criminal offenders. The results from using samples from Australia, Canada, England, Singapore, and two samples from the United States (North Carolina and Pennsylvania) indicated that (a) the SAQ has sound psychometric properties, with acceptable reliability and concurrent validity for assessing recidivism and institutional adjustment; (b) there were nosignificant differences among the scores of the White, African American, Hispanic, and Aboriginal Australian offenders on the SAQ; (c) there were no significant differences among offenders who completed the SAQ for research purposes versus offenders who completed it as part of a decision-making process. Results provided support for the validity of the SAQ to be used with the culturally diverse offenders involved in this research and provided further evidence that contradicts concerns that the SAQ as a self-report measure may be susceptible to lying, and self-presentation biases.

The aim of this study was to determine whether the Self-Appraisal Questionnaire (SAQ), a tool that was found to be reliable and valid for assessing violent and nonviolent recidivism and institutional adjustment for Canadian offenders, would also be valid for the same purposes with a demographically different population of North Carolina offenders. The internal consistency alphas and SAQ total and subscale scores' correlations were high. Offenders with high SAQ total scores had significantly more violent offenses, had more total number of past offenses, had higher numbers of past arrests, and had more institutional infractions than those with low SAQ scores. There were no significant differences between the responses of the African American and Caucasian offenders on the SAQ scales. These results support previous findings regarding the reliability and validity of the SAQ for assessing recidivism and institutional adjustment and suggest that the SAQ could be used with diverse populations.

The thalamic connectivity and basal forebrain cholinergic input to the posterior parietal cortex (PPC) of Long-Evans rats was examined using combined retrograde tracing and immunocytochemical methods. As in previous studies, the PPC could be distinguished by its input from the lateral posterior, lateral dorsal, and posterior nuclei of the thalamus, but not the lateral geniculate nucleus or ventrobasal complex. These nuclei were also observed to receive reciprocal projections from the ipsilateral PPC. Cholinergic neurons innervating the PPC were primarily localized to the substantia innominata/nucleus basalis region. The implications of these data for possible functions of the cholinergic input to PPC are discussed.

Considerable evidence suggests that various discrete nuclei within the amygdala complex are critically involved in the assignment of emotional significance or value to events through associative learning. Much of this evidence comes from aversive conditioning procedures. For example, lesions of either basolateral amygdala (ABL) or the central nucleus (CN) interfere with the acquisition or expression of conditioned fear. The present study examined the effects of selective neurotoxic lesions of either ABL or CN on the acquisition of positive incentive value by a conditioned stimulus (CS) with two appetitive Pavlovian conditioning procedures. In second-order conditioning experiments, rats first received light-food pairings intended to endow the light with reinforcing power. The acquired reinforcing power of the light was then measured by examining its ability to serve as a reinforcer for second-order conditioning of a tone when tone-light pairings were given in the absence of food. Acquisition of second-order conditioning was impaired in rats with ABL lesions but not in rats with CN lesions. In reinforcer devaluation procedures, conditioned responding of rats with ABL lesions was insensitive to postconditioning changes in the value of the reinforcer, whereas rats with CN lesions, like normal rats, were able to spontaneously adjust their CRs to the current value of the reinforcer. The results of both test procedures indicate that ABL, but not CN, is part of a system involved in CSs' acquisition of positive incentive value. Together with evidence that identifies a role for CN in certain changes in attentional processing of CSs in conditioning, these results suggest that separate amygdala subsystems contribute to a variety of processes inherent in associative learning.

This is the first of two papers describing the organization and connections of the ventral lateral geniculate complex (GLv) in the tree shrew. Using a combination of Nissl, Golgi, histochemical, and immunocytochemical methods, we have identified two major divisions (lateral and medial) of GLv, both of which can be further subdivided. The lateral division contains three subdivisions, external, internal and intergeniculate leaflet. The medial division contains two subdivisions, medio-rostral and medio-caudal. All three lateral subdivisions receive input from the retina, the densest terminations being in the external subdivision and intergeniculate leaflet. These projections originate primarily from small retinal ganglion cells, although a few large retinal ganglion cells also project to GLv by way of collateral branches. Each subdivision of GLv has a distinct cytoarchitectonic and immunocytochemical make-up. In general, the level of immunoreactive endings for glutamic acid decarboxylase (GAD), leuenkephalin (ENK), and choline acetyltransferase (ChAT) parallels the distribution of retinal projections. Thus, all three markers are particularly dense in the external subdivision and the intergeniculate leaflet. Cell bodies immunoreactive for ENK are restricted to the external and intergeniculate leaflet subdivisions. The medial subdivisions stain relatively poorly for GAD, ENK, and ChAT, although each has other cytological features that differentiate them from the lateral subdivisions and the adjacent thalamic reticular nucleus.

Connections of the ventral lateral geniculate complex (GLv) in the tree shrew were traced by anterograde and retrograde transport of WGA-HRP. The results buttress earlier findings that GLv in this species is composed of two main divisions, lateral and medial, each of which differs in its connections with the brainstem and cerebral cortex. The connections of the lateral division (GLv) suggest that it participates in visuosensory functions: it receives input from the retina, striate cortex, pretectum, and retino-recipient layers of the superior colliculus. These connections help clarify the identification of the internal and external subdivisions of GLv inasmuch as projections from both the superior colliculus and pretectum terminate in the external subdivision and each, in turn, receives a projection from the internal subdivision.Connections of the medial division suggest that this part of the nucleus is involved with visuomotor functions. Thus, the medio-caudal subdivision projects to the pontine nuclei, the prerubral field and the central lateral nucleus. The medio-caudal subdivision also receives projections from the lateral cerebellar nucleus, so that the GLv-ponto-cerebello-GLv loop involves mainly one subdivision of GLv. The medio-rostral subdivision receives projections from the pretectum and parietal cortex. Its output is directed primarily at the intermediate and deep layers of the superior colliculus. All of these targets of GLv, the pons, prerubral field, and deep layers of the superior colliculus, are known to play a role in the coordination of head and eye movements.Additional connections of GLv with the vestibular nuclei, intralaminar nuclei, hypothalamus, and facial motor nucleus are also described.

In this report we examine the dendritic organization of putative interneurons (class II cells) in different layers of the dorsal lateral geniculate nucleus of the tree shrew. The results show that there is considerable morphological diversity within this class, but that two broad groups can be identified: neurons whose dendrites remain within a layer or its adjacent interlaminar zones (intralaminar class II cells); and neurons whose dendrites cross into an adjacent layer(s) (interlaminar class II cells). The majority of class II cells in every layer have intralaminar dendrites, some of which are oriented along a particular axis, and others that are organized radially. The paired layers (1 and 2, 4 and 5) contain a particular group of intralaminar class II cells that have radially organized dendrites and elaborate claw-like appendages. The dendrites of interlaminar class II cells are organized along lines of projection and extend across as many as four layers. These cells often reside close to or within the interlaminar zones. Overall, the organization of class II cells seems to follow a pattern similar to the class I (relay) cells identified previously. Most have intralaminar dendrites, which presumably underlie the fidelity of signals transmitted from the retina to a particular layer. However, there are also a number of other cells whose processes cross laminar borders, presumably to affect integrative functions within the nucleus.

In this study we examined the organization of projections from the striate cortex to the dorsal lateral geniculate (GL) and pulvinar (PUL) nuclei in the prosimian Galago by using retrograde transport methods. Injections of wheat germ agglutinin-conjugated horseradish peroxidase (WGA-HRP) into the PUL labeled two bands of cells in the striate cortex: the first consisted of large pyramidal cells in the upper half of layer V; the second consisted of small and medium-size pyramidal cells located in the deepest part of layer VI. The location of cells within layer VI coincided with a clear cytoarchitectonic sublayer, VIb, which contains fewer and paler staining cells than VIa. Injections of WGA-HRP involving all layers of the GL produced an uninterrupted band of pyramidal cells distributed throughout layer VI (a and b), including the region labeled after injections into the PUL. Thus as a first approximation, layer VI can be divided into an upper tier (VIa) that projects only to the GL and a lower tier (VIb) that projects to both the GL and PUL. Injections of WGA-HRP that were restricted to one or a few GL layers revealed a further refinement of the subdivisions within layer VI. Injections into the parvicellular and intercalated (or koniocellular) layers of the GL labeled neurons predominantly in the upper half of layer VIa, whereas injections restricted to the magnocellular layers labeled neurons in the lower half of layer VIa and in layer VIb. In order to determine whether individual neurons in layer VIb send axon collaterals to both the GL and PUL, we injected WGA-HRP into one nucleus and fluorescent rhodamine latex beads into the other. In three experiments, we found only one double-labeled cell. In sum, the results provide evidence that layer VI is divided into at least three sublayers: upper VIa, which projects to the intercalated and parvicellular GL layers; lower VIa, which projects to the magnocellular GL layers; and VIb, which sends separate projections to the magnocellular layers of the GL and to the PUL. The segregation observed is sufficiently discrete to propose the existence of multiple, descending pathways from layer VI of the striate cortex that complement those ascending from the GL and PUL.

The types of laminar segregation of neuronal classes found in the lateral geniculate nucleus (GL) of mammals are almost as varied as the groups themselves which exhibit lamination. This essay deals with the question of whether, in spite of the differences in which classes are segregated by layer in different species, one rule might apply to the laminar distribution of projections from the GL to the striate cortex. Until recently, the weight of experimental evidence suggested that the laminar or sublaminar distribution of such projections was organized so as to segregate functional classes (i.e. W, X, Y, on center, off center). While this model appears to be consistent with the organization of geniculostriate projections in many species (especially primates), recent studies on species other than primates suggest that the laminar distribution of geniculostriate projections is not based on a segregation of functional classes per se, but on the segregation of projections from whole GL layers (or layer pairs) including all of their constituent cell classes.

The neuronal organization of the lateral geniculate nucleus of the prosimian primate, Galago crassicaudatus , was studied in Golgi‐Kopsch‐impregnated material. On the basis of cytoarchitecture, electrophysiology, and connections the nucleus is divisible into three pairs of layers—one magnocellular, one parvocellular, and one koniocellular—each part of a separate retinogeniculate and geniculostriate pathway (Itoh et al., '82; Norton and Casagrande, '82). In Macaca and Saimiri , which have equally distinct geniculate subdivisions, it has been reported that, outside of cell size, no one morphological attribute differentiates magnocellular from parvocellular neurons (Campo‐Ortega et al., '68; Wong‐Riley, '72; Saini and Garey, '81; Wilson and Hendrickson, '81). Results presented here are not inconsistent with this conclusion. However, when the results are analyzed from the standpoint of the collective traits that distinguish the cell groups that make up the layers, clear morphological differences are evident. Using this approach we find the following differences between presumed projection neurons and interneurons in each pair of layers. The projection neurons of the magnocellular layers, as a group, exhibit large cell bodies with radially arranged dendrites which often extend beyond laminar borders. The magnocellular interneurons are larger than their counterparts in the other layers and, like the magnocellular projection neurons, exhibit radially arranged dendrites. The former, however, also share characteristics in common with other interneurons such as relatively small somata, few proximal dendrites, and complex distal dendritic appendages. In contrast, the projection neurons and interneurons of the parvocellular layers have smaller somata and more restricted dendritic spreads than their counterparts in the magnocellular layers. Dendritic arbors of parvocellular neurons are typically oriented perpendicular to laminar borders and remain confined to their layer of origin. The koniocellular neurons represent a more diverse population but collectively are distinct in that the dendrites of almost all neurons in these layers run parallel to the layers. The fact that presumed interneurons and projection neurons in a single layer share a number of related dendritic features suggests that both groups together are responsible for the structural and, hence, functional architecture of a layer. We conclude that when the organization of the lateral geniculate nucleus is considered from the standpoint of layers and not individual cell types, the functional differences between layers indicated bydifferential connections, histochemistry, and electrophysiology can be related to distinct morphological differences singular to magnocellular, parvocellular, and koniocellular divisions.

The genus Galago provides an unique opportunity to study the relation between layers of the lateral geniculate body and classes of retinal ganglion cells. In the present experiments HRP was restricted to individual layers of the lateral geniculate body with the following results: After injections of the magnocellular layers, layers 1 and 2, labeled retinal ganglion cells ranged in size from 8 to 20 μm. After injections of the parvocellular layers, layers 3 and 6, labeled retinal ganglion cells ranged in size from 6 to 12 μm. After injections involving layers 4 and 5, which layers contain only very small, pale cells, labeled retinal ganglion cells ranged in size from 5 to 14 μm. Thus, the very largest ganglion cells were labeled only after injections of magnocellular layers 1 and 2, while small and medium retinal ganglion cells were labeled after HRP injections in every layer of the lateral geniculate body. Because the magnocellular layers actually contain a mixture of large, medium, and small‐sized cells, we suggest that retinal ganglion cells of different size‐classes project to geniculate relay cells of the corresponding size‐class.

Using both binocular and monocular viewing conditions, optokinetic nystagmus (OKN) frequency-velocity functions were measured before and after ophthalmic laser-produced retinal lesions. With binocular viewing conditions, the range of effective pattern velocities and the upper velocity threshold increased by 20–30°/sec in subjects with both foveas lesioned, and by 12°/sec in a subject with only a single foveal lesion. Subjects with parafoveal lesions showed no change in postlesion binocular OKN response functions.Prelesion monocular OKN functions were obtained for both temporal-to-nasal (T-N) and nasal-to-temporal (N-T) directions of pattern movement. T-N OKN functions were similar to those obtained with binocular viewing but N-T movement elicited OKN over a much narrower range of stimulus velocities. Lesions of the foveal area appeared to have little effect on N-T OKN functions, although an increase in N-T upper velocity thresholds was obtained from at least one eye in 4 of 5 animals. Thus, lesions of the foveal and parafoveal area do not impair, and may actually facilitate, OKN in the pigeon.