Trading Consistency for Scalability in Scientific Metadata

Size: px
Start display at page:

Download "Trading Consistency for Scalability in Scientific Metadata"

Transcription

1 Trading Consistency for Scalability in Scientific Metadata Scott Jensen and Beth Plale School of Informatics and Computing Indiana University Bloomington, Indiana Abstract Long-term repositories that are able to represent the detailed descriptive metadata of scientific data have been recognized as key to both data reuse and preservation of the initial investment in generating the data. Detailed metadata captured during scientific investigation not only enables the efficient discovery of relevant data sets but also is a source for exploring ongoing activity. In XMC Cat metadata catalog, an XML catalog that uses a novel hybrid model to store XML to a relational database, we exploit differences in the temporal utility between browse and search metadata to selectively relax the consistency model used. By ensuring only eventual consistency on parts of the solution, we determine through experimental analysis that the performance and scalability of the catalog can be substantially improved. Keywords-Scientific Metadata, Eventual Consistency, XML, e- Science I. INTRODUCTION In e-science, metadata repositories have been recognized as crucial to preserving an organization s investment in data by enabling metadata collection, long-term preservation, and reuse of scientific data. This includes both observational data and data sets derived through scientific investigation and analysis. Science communities facilitate communication within their community through development of domain-specific XML schemas often as an extension/profile of metadata standards such as the FGDC [12] and ISO [17] for geospatial data. Researchers in computer science and library science have categorized the metadata captured by these detailed schemata [16,25,23,31,33]. Common to such categorizations is a breakdown between discovery and administrative or structural metadata; where discovery metadata is used by researchers to locate datasets of interest, and administrative or structural metadata is needed to actually use the dataset once it has been discovered [16]. With the XMC Cat metadata catalog developed in the context of the Linked Environments for Atmospheric Discovery (LEAD) project [9], our focus is on automating the capture of detailed discovery metadata during scientific investigation utilizing a workflow-oriented cyberinfrastructure. The LEAD science gateway currently has 126 users, with 5TB of atmospheric data, workflows, and forecast results. This collection is described by over 2 million metadata concepts containing more than 8 million metadata elements. Our focus in automating the capture of metadata is two-fold. As noted by Gray et al., metadata is ephemeral and must be captured as data This research was funded in part by NSF CNS is generated since it cannot be recreated later [15], so capture during workflow execution enables long-term preservation by ensuring that ephemeral metadata is not lost. However, it also addresses a recognized issue in archiving scientific data the benefits of metadata annotation and archiving inure to the organization and not the scientist generating the data [6,36]. Metadata captured during workflow execution provides value to scientists as a tool to monitor and review results prior to publication. In a science discovery cyberinfrastructure such as LEAD four major activities determine the workload on the metadata catalog (Fig. 1): browse, search, catalog new data objects, and automated harvest of metadata. Through a portal interface a researcher can browse detailed metadata in a time-organized hierarchy of their experiments, including inputs, intermediate results, and workflow output, enabling a researcher to monitor ongoing, long-running experiments. In this context, only the object s ID is needed to retrieve the metadata that describes it. A search interface allows scientists to search both private and public data based on detailed metadata captured for each experiment; including all configuration settings and critical workflow notifications. As experiments are executed, core metadata is captured about inputs, the experiment itself, intermediate data products, and the output generated. Figure 1. Four operations dominate a metadata catalog s workload The dotted lines in Fig. 1 between harvest and browse illustrates a relationship that is not applicable to long-term archives concurrent with ingest, metadata is used by scientists to monitor workflow progress. If a workflow has bursts of data generation activity, this translates into a workload for the metadata catalog that is bursty in nature. Analyzing the metadata catalog operations of a 15-hour weather forecast workflow for instance, the workflow takes 1 hour to run from initializing boundary conditions through to generating 2D visualizations that are viewable on a cell phone [29]. Although the entire workflow takes 1 hour to execute, the metadata ingest operations are highly non-uniform throughout

2 the hour. When the workflow is first launched, metadata for the experiment and its collections are inserted within the first minute of the workflow. This insert includes core metadata, the workflow, and initial configuration parameters. Once the job has moved off the wait queue of the supercomputer, all forecast model initialization data are added in a burst. After this initial spin-up, the early stages of model initialization and data assimilation execute relatively quickly, resulting in a handful of early workflow stages completing within 30 minutes and cataloging their metadata, followed by a 20 minute lull (other than periodic workflow notifications) as the main forecast model (WRF [24]) completes. Were one to catalog metadata in a long-term archive, the bursty behavior of workflows would not be relevant since the workflow run had finished executing long before results of the experiment are published. Even if submitted in bursts postmortem, the inserts can be queued to balance the load. The bursty nature of metadata workload created by scientific workflows could be smoothed by more hardware such as through replication or clustering, but aside from the hardware cost, this also imposes a cost on browse queries which depend on concurrently ingested metadata. We show that the scalability of cataloging scientific metadata in response to an increased workload can be addressed through loosening the consistency required between browse and search metadata by exploiting temporal differences in when metadata is needed for these two operations. The remainder of this paper is organized as follows: Section 2 discusses eventual consistency. Section 3 discusses the application of eventual consistency to the use of metadata in browsing and searching scientific metadata. Section 4 gives details of an experimental evaluation of eventual consistency as applied to scientific metadata. Section 5 discusses related work, and section 6 concludes with future directions. II. EVENTUAL CONSISTENCY IN DATA ACCESS In relational databases, changes to the data are commonly performed through transactions that ensure the properties of Atomicity, Consistency, Independence, and Durability (ACID) [14], that is, the set of changes comprising a transaction are either performed in their entirety or not at all (atomic), the set of changes does not interact with other transactions and are thus serializable (independent), the changes are permanent once the transaction completes (durable), and the constraints imposed on relationships are maintained (consistency). With the emergence of the Internet, cloud-based web services, and large-scale data stores, the need for scalability and availability raises a challenge to the feasibility of ACID properties. One of the earliest discussions on the trade-offs between ACID and availability is by Eric Brewer, co-founder of Inktomi, who at a 2000 PODC keynote introduced Basically Available Soft-state Eventual consistency (BASE) [8]. Proposed as an alternative to ACID, BASE acknowledges the tension between consistency and availability. Brewer states that there are three properties of distributed systems (1) consistency, (2) availability, and (3) tolerance of network partitions, and that only two of these are possible at one time a property he refers to as the CAP theorem. As noted in Vogels [35], partitions are a given for large distributed systems, so under the CAP theorem, consistency must be traded off to gain availability. However, Brewer views ACID and BASE not as two absolutes, but as forming a spectrum with ACID providing strong consistency and BASE providing weak consistency in which approximate answers and stale data are accepted. In application of CAP to management of scientific metadata, scalability in the face of bursty workloads can similarly be addressed by relaxing consistency in order to provide fresh results where fresh results are needed, and allowing some operations to see older data where time critical needs are not as great. In our research, this tradeoff comes in preserving freshness for browse operations while at the same time allowing the search metadata to gain consistency over a longer period of time, that is, eventually. III. E-SCIENCE DISCOVERY METADATA A. Browse and Search Metadata as Functional Groups Capture of discovery metadata is for two distinct uses, browsing and searching, and the temporal utility of these two types of metadata differ. Browsing is most useful when the data of interest is known in advance, such as during workflow execution or reviewing a set of search results. In contrast, search capabilities address the longer-term preservation of a community s investment in data by allowing a scientist to find datasets that address her needs based on querying for measured variables, configuration settings, data quality metrics, or other detailed discovery metadata. In contrast to browse metadata which can be located based solely on the unique ID of a data object, search metadata requires a greater level of detail to be of value as data ages corresponding as Michener notes with our lack of ability to hold such details in memory [25]. In the XMC Cat metadata catalog, the metadata used to reconstruct XML in response to browse queries is managed separate from the detailed shredded metadata used to search for data products. This hybrid XML-relational approach [19] is used to maximize the performance of browse queries that are executed as users interactively click on data products in the web portal. It is this bifurcation of browse and search representations of the metadata that is exploited to relax the consistency between these two subclasses of discovery metadata. As Pritchett notes in [30], consistency is always in a state of flux when it is relaxed, and since temporal inconsistency cannot be hidden from users, one must look for opportunities where consistency can be relaxed preferably between different functions in a system. Browsing and searching can be viewed as two distinct functions. We exploit this distinction to improve the scalability of the metadata catalog and its adaptability under bursty workloads. B. Introduction to XMC Cat XMC Cat is a Java based web service metadata catalog that uses a lightweight software stack based on Apache Tomcat and the Axis2 web services engine [3]. It adapts to community metadata schemata by partitioning each schema based on the metadata concepts it contains [20]. When ingesting metadata, character large objects (CLOBs) containing the XML for each concept are stored in the catalog and later used to rebuild the

3 full metadata for an object in response to queries. Browse queries require only these CLOBs to build a response. In addition, each concept is also parsed using a process commonly referred to as shredding [32]. The shredded elements and the concept structure containing them provide the detailed discovery metadata needed for search queries. If the CLOBs and shredded metadata for a concept are stored using a single transaction, the recall and precision of the search results depends only on the accuracy and completeness of the metadata cataloged. Any object that matches the search criteria would be found (recall) and any object listed in the results would match the search criteria (precision). Ensuring consistency between browse and search metadata requires the CLOBs and shredded metadata to be inserted as an atomic operation. The shredding process transforms metadata based on a domain schema to a domain-neutral concept and element structure using the following transformations: Metadata Document Globally Unique ID Concept + Concept Global order (int) CLOB Shredded Concept Shredded Concept Concept name (string) Concept source (string) Shredded Concept * Shredded Element * Shredded Element Element name (string) Element source (string) Element value (string) During ingest, each concept contained in the metadata is tagged with its global order based on the partitioning of the schema. The CLOBs are indexed based on the object ID and global order of the concept they represent. This covering index is used to determine the concepts that describe an object; allowing metadata to be quickly reconstructed in response to queries even if it was added incrementally through additional metadata harvesting. C. Eventual Consistency of Search Metadata An earlier performance evaluation of XMC Cat s hybrid XML-relational approach compared it to in-lining (an alternative often used for storing XML in a relational database), and found shredding and storing the leaf elements to be a significant cost [18]. Storing concepts as both CLOBs and shredded elements greatly improves query response time, so shredding cannot simply be done away with. However, since the CLOBs and shredded elements are used for different functions (browsing versus searching), both insert performance and scalability can be improved by relaxing the consistency between browse and search metadata and only ensuring eventual consistency instead. Eventual consistency requires changes to both the metadata catalog s architecture and the ingest process as shown in Fig. 2. Eventual consistency is implemented architecturally by the introduction of concept shredders deployed on distributed servers which asynchronously shred concepts into searchable metadata. Under this new approach, only a subset of the first two transformations is performed during the ingest process - Storage of concepts so a user can browse their workspace Distributed shredding of concepts for eventually consistent quering of the workspace XMC Cat Web Service parse metadata into concept-sized CLOBs 1a 1b store concept CLOBs to object s metadata shredded metadata added to XMC Cat metadata catalog 6a adding metadata to existing experiments XMC Cat shred into sub-concepts and elements Yes Metadata Catalog successfully shredded? new experiments adding metadata queue concept ID for eventual shredding queue query for a batch of concept CLOBS to process 2 Distributed Concept Shredders remove CLOBs added to queue 3 entry for CLOB in processing dequeue CLOB 4 queue worker thread... worker thread 6b Distributed Concept Shredders Figure 2. Distributed shredding of metadata concepts. The shredding of search metadata is queued and performed asynchronously. parsing out the concept CLOBs (Step1a in Fig. 2). Additionally, the internal IDs of the CLOBs are inserted into a FIFO queue for eventual processing (Step 1b). These steps are captured by the following reduced set of transformations: Metadata Document Globally Unique ID Concept + Concept Global order (int) CLOB Under this approach, the shredders are distributed to different servers and independent no shredder is aware of other shredders, and the XMC Cat catalog does not know how many shredders are operational at any point in time. Each distributed shredder has a controller thread that fetches batches of concepts (the CLOB, its ID, and version) from the catalog s queue (Step 2) and stores them in its own local queue (Step 3). The frequency of retrieval adjusts automatically based on the number of concepts remaining in the local queue at the start of each fetch cycle. Although fetched by a shredder for processing, to prevent the loss of queryable metadata should a shredder fail, concepts must remain in the catalog s queue until the CLOB is actually shredded and inserted back into the catalog. To minimize the retrieval of concepts by multiple shredders, an approach similar to that used in [7] is employed. Each concept in the catalog s queue has a timestamp that is set on retrieval to a configurable time forward when it can again be retrieved. Although the queue is a FIFO queue, concepts are not retrieved by other shredders until their timestamp has expired. A set of worker threads at each shredder dequeue concepts from the local queue in a synchronized manner (Step 4) and performs the remaining transformations for shredding the CLOBs into a generalized concept and element Yes 5

4 representation (Step 5) and inserting the shredded metadata into the catalog (Step 6a). After the shredded metadata is inserted, the entry in the catalog s concept queue can be removed (Step 6b). Idempotent Concept Inserts: Inserting the shredded metadata into the catalog in Step 6a must be idempotent since multiple inserts could otherwise occur in the following two situations: 1. If an existing concept is updated, the prior version and the update could both be in the queue to be shredded. If the controller at each distributed shredder retrieves a large batch of concepts to reduce frequent fetching, the update could be processed by one shredder (because it is near the front of its local queue) before the prior value was shredded. Each CLOB has a version assigned and if the version for the update does not match the current CLOB, the update is abandoned. 2. If a shredder is slow, it may not complete processing a concept before the future timestamp set during the fetch cycle in Step 2 expires. This could result in two distributed shredders fetching the same concept. In this case, a unique key in the shredded concept table prevents the insert of an identical concept for the same object and CLOB ID, so the insert is abandoned. D. Bounds on Eventual Consistency for Search Metadata The argument for using eventual consistency in managing scientific metadata is improved insert performance and scalability. However, if the time lapse until consistency is achieved exceeds the temporal difference in the use of browse and search metadata, search results will suffer. The lower and upper bounds on achieving consistency is defined as ECt and determined as follows: EC t = W t + T t + R t + S t + I t Where W t, T t, and R t represent the queue and fetch time from when a concept is first queued at the catalog in Step 1b until it has been loaded into a shredder s local queue in Step 3. W t is the time a concept waits in the catalog s queue after being queued in Step 1b until it is retrieved by a shredder in Step 2. The fetching of a concept and loading it into a shredder s local queue (Steps 2 and 3) can be decomposed into two operations, the time to identify or tag the concepts to be retrieved (T t ), and the time to retrieve and load the tagged concepts (R t ). The total time for a distributed shredder to process a concept (Steps 4 6) is the sum of S t and I t where S t is the time required to shred the concept (Steps 4 and 5) and I t is the cost of inserting the shredded metadata into the catalog (Step 6). W t depends on two factors, the time the distributed shredder s controller waits to fetch (C t ) and whether the insert rate into the concept queue in Step 1a exceeds the maximum capacity service rate of the distributed shredders. When the shredders are processing on par with the insert rate, W t = C t (which is configurable for each distributed shredder). The default lower and upper bounds on C t are 1 second and 120 seconds respectively. The lower bound for EC t is dominated by C t since if a concept is queued midway through the minimum wait time of 1,000ms, the wait would be 500ms. We determine by means of experimental investigation the average component costs to fetch and process concepts. The average time to fetch a batch of 100 concepts and load the shredder s local queue is 139ms, which breaks down almost evenly between 64.42ms on average to tag the concepts (T t ) and 74.58ms to retrieve and load the concepts (R t ). The time to process the concepts averages 17.22ms, with the time to shred averaging 3.48ms and the time to insert the shredded metadata averaging 13.74ms. As with the lower bound, the average time delay for eventual consistency is also dominated by the wait time in the queue (W t ) since the average time for fetching, processing, and inserting a concept totals only ms. Since the insertion of each CLOB s ID into the concept queue in Step 1b is performed in a stored procedure that is executed atomically with the insert of the CLOBs in Step 1a, no delay in consistency occurs to initially queue the concept. IV. EXPERIMENTAL EVALUATION A. Insert Metadata Although XMC Cat supports a number of administrative operations, the main functionality of a catalog is inserting and querying metadata. To evaluate the benefits of eventual consistency, the following measurement metrics are used: Insert operation response time Scalability performance of inserts and queries For all tests we use real metadata harvested from the workspaces of meteorological researchers using the LEAD portal. The base scalability workload used in testing is adopted from [28], a 2007 study that estimated and characterized the volume of experiments that would be run in a production LEAD environment. Execution time of experiments generating metadata in LEAD is dominated by the execution time of the Weather Research and Forecasting (WRF) model. Likewise, the number of files generated by an experiment is also based on WRF. The metadata added for each experiment consists of spatial and temporal metadata, citation and usage elements from the FGDC schema that are similar to elements from the Dublin Core standard [10], notifications harvested from the LEAD notification bus during WRF experiments, and configuration parameters (contained in FORTRAN namelist configuration files). The metadata cataloged for files within an experiment are taken from the ARPS Data Analysis System (ADAS) files used in LEAD. The metadata for an individual file is less voluminous than for an experiment and consists of metadata describing the data s owner, distribution, spatial and temporal bounds, and terms from the Climate and Forecasting controlled vocabulary [11]. Each experiment catalogs 2,202 separate metadata elements grouped into 256 concepts which frequently exhibit some complex structure such as is the case for workflow notifications. The metadata catalog is initialized from the workspaces of 125 actual users. B. Test Environment The metadata catalog s web service is hosted on a Dell PowerEdge 6950 configured with quad dual-core AMD

5 Mean Execution Time (ms) Validate Request 2,500 Validate Parent XSLT XML Bean Root Element Define Object Extract Concepts Insert Concepts Shredded Metadata 2,000 1,500 1, Mean Execution Time (ms) Mean Execution Time (ms) Validate Request Validate Against Existing Metadata Insert Concepts Shredded Metadata Other 0 Eventual Strict (a) File Metadata Other 0 Eventual Strict (b) Experiment Metadata 0 Eventual Strict (c) Incremental Concepts Figure 3. Baseline comparing strict consistency with eventual consistency when inserting metadata. The performance improvement from eventual consistency is mainly due to inserting shredded search metadata asynchronously, but it also reduces the cost of parsing and manipulating the metadata using XML beans. 2.4GHz Opteron processors, 16GB of memory, and running RHEL 4. The web service stack uses version 1.3 of Apache Axis2 on Tomcat version The database is MySQL version The clients and distributed shredder are located on separate servers configured with dual 2-core 2.0GHZ AMD Opteron processors, 16GB of memory, running RHEL 4, and connected by Gigabit Ethernet. C. Baseline Performance Since browse metadata is of value to scientists in monitoring experiments, metadata describing a digital object must be able to grow incrementally during long-running workflows. This requires cataloging metadata for new objects and incrementally adding metadata for existing digital objects. To measure insert performance, we focus on these two operations and vary the size of the insert from moderately-sized metadata describing data files to the more substantial metadata used to describe WRF experiments. Additionally, we evaluate incremental cataloging of metadata for an experiment by initially inserting only core metadata and then incrementally adding metadata concepts such as notifications and configuration settings that would be added during workflow execution. XMC Cat also supports updates and deletions of existing metadata concepts for an object, but in LEAD those operations are much less frequently invoked. Metadata for both files and experiments are inserted for 25 iterations, and eliminating the spin-up of the first operation, Fig. 3(a) and 3(b) illustrate the respective mean execution time broken down into its component costs. Although it takes longer to insert the substantially greater metadata describing an experiment, the dominant costs for both files and experiments are (1) transforming domain metadata to concepts and elements using XSLT, (2) parsing the transformed metadata as an XML Bean, and (3) inserting the shredded metadata. For experiments, operations that manipulate larger XML documents start to become more expensive such as extracting the root metadata element from the shredded metadata. The majority of the savings in execution time under eventual consistency is the cost of inserting shredded metadata since that step is delegated to the distributed shredders and done asynchronously outside of the ingest operation. However, there is also a reduction in the time required for the transformation since the XSLT template and resulting XML bean are smaller due to delegating shredding of leaf elements to the distributed shredders. Overall strict consistency takes 41% and 58% longer respectively for files and experiments to execute the metadata insert operation in comparison to eventual consistency. Since metadata is added incrementally to experiments during workflow execution, we test the performance of adding concepts to previously cataloged experiments. As a baseline we ran 25 iterations that added 76 concepts containing actual workflow notifications and configuration parameters from WRF experiments in LEAD. Each concept is added in a separate web service call to XMC Cat. Fig. 3(c) shows the mean execution time excluding the first concept added. Consistent with files and experiments, the time saved under eventual consistency is mainly attributable to delegating inserts to the distributed shredders. However, the additional cost for strict consistency is only 26% because a significant portion of the time is used to check that adding the concept does not result in metadata that no longer validates against the schema. D. Catalog Workload The scalability evaluation is carried out using a pseudorealistic workload derived from usage patterns observed in the LEAD science gateway. The data and data characteristics of each of the experiments are realistic, taken with permission from the workspaces of users of the LEAD portal.plale [28] reports the characteristics of a workload based on 125 active users who each execute 4 workflows in a 12-hour time frame for a total of 500 workflows. These workflows are classified based on the following four categories with their respective frequencies: Educational Experiments (50%) Canonical Experiments (10%) Ensemble Experiments (1%) Data Import (39%) The number of files generated by a real workflow will vary depending on factors such as configuration. The workload estimates are based on average observed behavior. An ensemble experiment is set at 100 canonical experiments where

6 each canonical experiment generates 72 files. The educational experiments are smaller and generate only 16 files. The data import only adds files. Since metadata for experiments is added incrementally as the workflow executes, metadata for the canonical and ensemble experiments is added incrementally with the frequency of the incremental inserts based on the running time of the experiments and the number of concepts to be added. The metadata for each educational experiment is added as a single larger metadata insert. The query workload is based on a 76.9:23.1 total query-toinsert ratio. This is adopted from the TPC-E benchmark [34] because of the resemblance of usage captured in TPC-E and observed in the LEAD workspace. Based on the patterns observed in the LEAD portal, three quarters of the queries in the scalability test are classified as browse queries based on the object s unique ID and one quarter are search context queries. As a final aspect of the workload, a set of test clients execute each type of insert or query and are configured with a test plan that specifies different users for whom metadata should be added and queries executed. The test plan specifies the rate at which the client executes a query or insert, and each client executes the web service calls on a separate thread to prevent the event of saturation at the metadata catalog service from impeding the rate at which operations are launched. All scalability tests are run based on a 30-minute window. Specifically, after running the initial base case which is modeled on the projected LEAD workload, the workload is first doubled and then increased in subsequent runs by 2 times the base workload until the service is saturated. Each increment of the base workload adds 43,837 concepts containing 212,565 metadata elements that describe data files and experiments and can be queried by scientists to locate relevant datasets, so at 10 times the base workload, the scalability test is adding over 2 million queryable shredded metadata elements in the 30 minute window while executing a total of 62,827 browse and search queries. E. Scalability Performance The scalability piece of the evaluation compares total execution time of strict and eventual consistency under a workload that is increased in size through increments that are multiples of the base catalog workload described in sub-section D above. The turnaround time for inserts and queries are measured as the workload is scaled from 1 to 8 times the base catalog workload of inserts and queries. Query performance does not differ significantly except for higher overall scalability under eventual consistency. Insert performance is shown in Fig. 4 as the workload is scaled to 8 times the base workload. Fig. 4(a) shows the average time in milliseconds to insert metadata for a file as the workload is scaled from the initial base workload to 8 times that workload. The solid line indicates the total time for strict consistency and the dashed line represents the portion of that total spent inserting the shredded metadata. The dotted line shows the total time required under eventual consistency using distributed shredding. The performance difference is almost entirely attributable to inserting the shredded metadata (which under Mean Execution Time (ms) Mean Execution Time (ms) Including Validation Excluding Validation Figure 4. Scalability of inserts under an increased load. As the workload is scaled from 1 to 8 times the base workload, the performance of eventual consistency for files (a) scales well. Larger metadata inserts for experiments (b) are negatively impacted by validating the web service request. eventual consistency is distributed and done asynchronously). The advantage of eventual consistency begins to decrease at 6 times the base workload, but strict consistency still takes 42% longer to execute and even at that workload eventual consistency still performs better than strict consistency at the initial base workload. The dashed and dotted line in Fig. 4(a) shows the total time required for eventual consistency with the distributed shredder turned off. Under this alternative eventual consistency instead becomes more advantageous at 6 times the base workload since the cost of shredding under strict consistency begins to rise. Turning off the shredders increases the consistency delay between browse and search metadata, but if a gateway expects a bursty pattern, shredding can be adapted to this bursty behavior. When adding metadata concepts to existing experiments, eventual consistency performs only marginally better unless the shredders are turned off, in which case strict consistency takes 27% to 53% longer to execute. This is expected since determining the validity of the resulting metadata is a significant cost when incrementally adding concepts. The performance when inserting all of the metadata for an experiment as a single large document is depicted in Fig. 4(b). The total time to insert the metadata shows wide fluctuations in performance. The lower section of Fig. 4(b) shows the total cost for inserting the same metadata but excludes the initial validation of the request. Without this validation step, the performance fluctuations exhibited by both consistency models is eliminated and eventual consistency is advantageous. The non-linear cost of validating large metadata documents highlights a performance advantage to inserting metadata incrementally. Validating large XML Bean instances is CPUintensive and under the concurrent inserts and queries of the

7 scalability workload significantly degrades performance. In XMC Cat, metadata schemata are partitioned into concepts and validating concepts added incrementally does not exhibit this same performance cost. Small and relatively static tables identify and validate mutually exclusive and required concepts, so although concepts are added incrementally, this never results in an invalid metadata document. F. Moderating the Workload with Eventual Consistency As the workload scales up, eventual consistency allows the catalog to scale to 10 times the base workload without errors or timeouts that preclude successful completion of the web service calls. Under strict consistency the workload can only scale to 8 times the base workload without encountering fatal errors. While the difference may not appear great, hidden is the fact that under strict consistency non-fatal insert errors start to occur at 2 times the base workload. These non-fatal errors occur when the CLOBs have been inserted but shredding of the search metadata fails. For these inserts consistency is less wellbehaved than under eventual consistency. Eventual consistency provides the ability to automatically adjust the trade-off between consistency and scalability by dynamically adapting the configuration of the distributed shredders to moderate a heavy workload on the catalog. Under strict consistency, all search metadata must be shredded and inserted before an insert operation is complete. Alternately a non-fatal error occurs, but in our measurements handling these non-fatal errors does not shorten response time. Under eventual consistency, the number of shredders needed, and the number of worker threads at each shredder, provide an upper bound on the work performed to shred and insert search metadata in a given time frame. At 6 times the base workload the distributed shredders keep pace with the concept ingest rate. Of the 263,022 concepts added for that workload, only 100 remain in the queue at the end of the test. As the workload scales beyond this level the concept queue acts as a buffer. At 8 times the base workload the queue takes an additional 5 minutes to finish processing concepts. At 10 times the base workload the concept queue requires an additional 23.5 minutes to finish processing. The distributed shredder s controller thread automatically adjusts the interval between fetches to minimize the overhead incurred in calls to retrieve concepts. The fetch cycle is a two stage process where the manager first calls the catalog to determine if there are concepts available to process and tags them to be fetched. During this stage, if the catalog service is experiencing a heavy workload, a flag can be set with a timer to tell the shredder manager there are no concepts available. This allows the metadata catalog to dynamically adjust the number of active shredders based on the current system load. V. RELATED WORK Globus MCS [33] and SRB s MCAT [4] both take a domain-neutral approach to storing scientific metadata, and the EAV approach used in MCS for storing user defined attributes as name/value pairs (originally inspired by MCAT) influenced the storage of shredded metadata in XMC Cat. However, as noted by Singh et al. in [33], the conversion from XML to name/value pairs caused a performance bottleneck when storing metadata communicated as XML in MCS. In the ecological sciences, the Metacat metadata catalog [5,21] can ingest metadata as XML, shreds it, and stores it in a relational database. The approach used in Metacat to store XML is similar to an early schema-less approach referred to as the edge-table [13]. In Metacat, adding metadata incrementally to a previously cataloged object requires retrieving the entire XML document, updating it, and reinserting it. Reinserting the document also requires rebuilding all of the path indexes for the document making incremental updates expensive. Additionally, when an object s metadata is updated, the entire prior metadata for that object is archived [22], making hundreds of incremental concept inserts during workflow execution impractical. More recently, one of the motivating use-cases [2] for Amazon s SimpleDB was as a metadata index to data stored in Amazon s Simple Storage Service (S3). Although deployed in a cloud, SimpleDB uses an EAV approach similar to MCS and MCAT in that it uses a schema-less approach to store up to 256 simple name-value pairs per object. Unlike the shredded metadata stored in XMC Cat or the user-defined attributes stored in MCS, values in SimpleDB are not strongly typed but are instead stored as strings. Although, SimpleDB allows for range and value comparisons other than equality, numeric data must be consistently zero-padded since comparisons are lexicographical. SimpleDB does use eventual consistency for scalability and availability; updates are expected to propagate to all replicas within 2 seconds [1]. VI. CONCLUSION AND FUTURE WORK In this paper we have shown that the hybrid XML-relational approach used in the XMC Cat metadata catalog web service can exploit differences in the temporal utility of browse and search metadata to improve the performance of metadata inserts and overall system scalability. By relaxing the consistency between these two subcategories of discovery metadata to ensure only eventual consistency, the shredding of search metadata can be done asynchronously by delegating the task to distributed concept shredders. Evaluation of the catalog s ability to scale in the face of a incrementally expanding user workload shows that relaxing the consistency between browse and search metadata results in an almost 30% improvement for insert operations. This approach is also wellsuited to handling incremental updates of metadata communicated as XML allowing browse metadata to provide greater value to scientists as a tool for monitoring long-running experiments and workflows. Currently we are working with a research project in astronomy that is ingesting data based on a streaming model. In this context, we are exploring scalability and moderating of the catalog s workload in a streaming environment. The hybrid architecture in XMC Cat allows for a loose coupling between the community XML schema and the concepts scientists search over. Based on this loose coupling, we are researching the use of the metadata catalog, combined with semantic tools, as a front-end to legacy structured data, to

8 provide scientists with a query capability more closely aligned with their model of the data. The current implementation distributes the shredding of metadata concepts, but metadata is eventually inserted back into the main catalog. We are exploring approaches to distributing the metadata catalog itself to further scale with the deluge of scientific data being generated. ACKNOWLEDGMENT We thank the LEAD Gateway community, particularly the meteorological researchers at the University of Oklahoma and at other institutions using LEAD. We also thank the reviewers for their valuable feedback. REFERENCES [1] Amazon Web Services, Structured Data Storage, last accessed at: entry.jspa?externalid=3087&categoryid=152 [2] Amazon Web Services, Indexing Amazon S3 Object Metadata, last accessed at: [3] Apache Axis2 Next Generation Web Services. Available at: [4] C. Baru, R. Moore, A. Rajasekar, and M. Wan, The storage resource broker, in Proceedings of the 1998 Conference of the Centre for Advanced Studies on Collaborative Research (CASCON), Toronto, Canada, November [5] C. Berkley, M. Jones, J. Bojilova, and D. Higgins, Metacat: a schemaindependent XML database system, in Proceedings of the 13th International Conference on Scientific and Statistical Database Management, Fairfax, Virginia, July [6] J. Birnholtz and M. Bietz, Data at Work: Sharing in Science and Engineering, in Proceedings of the 2003 International ACM SIGGROUP Conference on Supporting Group Work, Sanibel Island, Florida, November [7] M. Brantner, D. Florescu, D. Graf, D. Kossmann, and T. Kraska, Building a database on S3, in Proceedings of the 2008 ACM SIGMOD International Conference on Management of Data,Vancouver, Canada, June [8] E.A. Brewer Towards robust distributed systems, keynote at the 19th ACM Principles of Distributed Computing (PODC 2000), Portland Oregon, July [9] K. Droegemeier, K. Brewster, M. Xue, D. Weber, D. Gannon, B. Plale, D. Reed, L. Ramakrishnan, J. Alameda, R. Wilhelmson, T. Baltzer, B. Domenico, D. Murray, A. Wlson, R. Clark, S. Yalda, S. Graves, R. Ramachandran, J. Rushing, and E. Joseph, "Service-oriented environments for dynamically interacting with mesoscale weather", Computing in Science and Engineering, IEEE Computer Society Press and American Institute of Physics, vol. 7, no. 6, pp , [10] Dublin Core Metadata Initiative, Available at: [11] B. Eaton, J. Gregory, B. Drach, K. Taylor, and S. Hankin, NetCDF Climate and Forecast (CF) Metadata Conventions, Version 1.0, [12] Federal Geographic Data Committee, Washington, D.C., Content Standard for Digital Geospatial Metadata Workbook Version 2.0, [13] D. Florescu and D. Kossmann, Storing and querying XML data using an RDBMS, IEEE Data Engineering Bulletin, vol. 22, no.3, pp , [14] H. Garcia-Molina, J.D. Ullman, and J. Widom, Database systems: the complete book, Prentice-Hall, Upper Saddle River, New Jersey, [15] J. Gray, A. S. Szalay, A. R. Thakar, C. Stoughton, J. vandenberg, Online scientific data curation, publication, and archiving, Microsoft Research, Technical Report MSR-TR , July [16] B. J. Hurley, J. Price-Wilkin, M. Proffitt, and H. Besser, The making of America II testbed project: A digital library service model (part III: implementing the MOA II service model), Council on Library and Information Resources, December 1999, last retrieved from: [17] International Organization for Standardization, Geographic Information Metadata (ISO19115:2003), [18] S. Jensen and B. Plale, Using characteristics of computational science schemas for workflow metadata management, In Proceedings of the 2008 IEEE Congress on Services, IEEE 2008 Second International Workshop on Scientific Workflows (SWF 2008), Hawaii, July [19] S. Jensen, B. Plale, S. Lee Pallickara and Y. Sun, A hybrid XMLrelational grid metadata catalog, Workshop on Web Services-based Grid Applications (WGSA'06) in association with International Conference on Parallel Processing (ICPP-06), August [20] S. Jensen and B. Plale, Extended abstract: Schema-independent and schema-friendly scientific metadata management Fourth IEEE International Conference on escience, pp , [21] M.B. Jones, C. Berkley, J. Bojilova, and M. Schildhauer, Managing Scientific Metadata, IEEE Internet Computing, vol. 5, no. 5, pp , Sept./Oct [22] Metacat Administrator s Guide for Metacat 1.9.1, version 1.0, April Available at: [23] METS Metadata Encoding & Transmission Standard: An Overview & Tutorial, last retrieved from [24] J. Michalakes, J. Dudhia, D. Gill, J. Klemp and W. Skamarock, Design of a Next-Generation Regional Weather Research and Forecast Model, Towards Teracomputing, pp , World Scientific, River Edge, New Jersey (1998). [25] W. K. Michener, J. W. Brunt, J. J. Helly, T. B. Kirchner, and S. G. Stafford, Nongeospatial metadata for the ecological sciences, Ecological Applications, vol. 7, no. 1, pp , February [26] National Information Standards Organization (NISO), Understanding metadata, NISO Press, last retrieved from: [27] S. Newhouse, J. M. Schopf, A. Richards, and M. P. Atkinson, Study of user priorities for e-infrastructure for e-research (SUPER), in Proceedings of the UK e-science All Hands Conference, [28] B. Plale, Workload characterization and analysis of storage and bandwidth needs of LEAD workspace, LEAD TR 001, Linked Environments for Atmospheric Discovery (LEAD) (2007) Version 3.0. [29] B. Plale, LEAD II/Trident workflows for timely weather products: the challenge of Vortex2, Microsoft External Research Symposium, April [30] D. Pritchett, BASE: An ACID alternative, ACM Queue, vol. 6, no. 3, pp , May [31] A. Rajasekar and R. Moore, Data and metadata collections for scientific applications, in Proceedings of the 9th International Conference on High-Performance Computing and Networking (HPCN Europe), June [32] M. Rys, D. Chamberlin, and D. Florescu, XML and relational database management systems: the inside story, in Proceedings of the 2005 ACM SIGMOD International Conference on Management of Data, New York, June [33] G. Singh, S. Bharathi, A. Chervenak, E. Deelman, C. Kesselman, M. Manohar, S. Patil, and L. Pearlman, A Metadata catalog service for data intensive applications, in Proceedings of the 2003 ACM/IEEE conference on Supercomputing, Phoenix, November [34] Transaction Processing Performance Council: TPC Benchmark E Standard Specification Version (2007). [35] W. Vogels, Eventually consistent, Communications of the ACM, vol. 52, no. 1, pp , January [36] A. Voss et al., e-research infrastructure development and community engagement, in Proceedings of the UK e-science All Hands Meeting 2007, Nottingham UK, September 2007.

Evaluation of Two XML Storage Approaches for Scientific Metadata

Evaluation of Two XML Storage Approaches for Scientific Metadata Evaluation of Two XML Storage Approaches for Scientific Metadata Scott Jensen, Devarshi Ghoshal, Beth Plale Indiana University scjensen@indiana.edu, dghoshal@indiana.edu, plale@indiana.edu Abstract Scientific

More information

Representing LEAD Experiments in a FEDORA digital repository

Representing LEAD Experiments in a FEDORA digital repository Representing LEAD Experiments in a FEDORA digital repository You-Wei Cheah, Beth Plale Indiana University Bloomington, IN {yocheah, plale}@cs.indiana.edu IU-CS TR666 ABSTRACT In this paper, we discuss

More information

Using Characteristics of Computational Science Schemas for Workflow Metadata Management *

Using Characteristics of Computational Science Schemas for Workflow Metadata Management * Using Characteristics of Computational Science Schemas for Workflow Metadata Management * Scott Jensen and Beth Plale Computer Science Department, Indiana University scjensen@cs.indiana.edu, plale@cs.indiana.edu

More information

Performance Evaluation of the Karma Provenance Framework for Scientific Workflows

Performance Evaluation of the Karma Provenance Framework for Scientific Workflows Performance Evaluation of the Karma Provenance Framework for Scientific Workflows Yogesh L. Simmhan, Beth Plale, Dennis Gannon, and Suresh Marru Indiana University, Computer Science Department, Bloomington,

More information

Schema-Independent and Schema-Friendly Scientific Metadata Management 1

Schema-Independent and Schema-Friendly Scientific Metadata Management 1 Schema-Independent and Schema-Friendly Scientific Metadata Management 1 Scott Jensen and Beth Plale Indiana University Department of Computer Science scjensen@cs.indiana.edu, plale@cs.indiana.edu Abstract

More information

10. Replication. Motivation

10. Replication. Motivation 10. Replication Page 1 10. Replication Motivation Reliable and high-performance computation on a single instance of a data object is prone to failure. Replicate data to overcome single points of failure

More information

Introduction to Grid Computing

Introduction to Grid Computing Milestone 2 Include the names of the papers You only have a page be selective about what you include Be specific; summarize the authors contributions, not just what the paper is about. You might be able

More information

HTRC Data API Performance Study

HTRC Data API Performance Study HTRC Data API Performance Study Yiming Sun, Beth Plale, Jiaan Zeng Amazon Indiana University Bloomington {plale, jiaazeng}@cs.indiana.edu Abstract HathiTrust Research Center (HTRC) allows users to access

More information

Database Assessment for PDMS

Database Assessment for PDMS Database Assessment for PDMS Abhishek Gaurav, Nayden Markatchev, Philip Rizk and Rob Simmonds Grid Research Centre, University of Calgary. http://grid.ucalgary.ca 1 Introduction This document describes

More information

CS848 Paper Presentation Building a Database on S3. Brantner, Florescu, Graf, Kossmann, Kraska SIGMOD 2008

CS848 Paper Presentation Building a Database on S3. Brantner, Florescu, Graf, Kossmann, Kraska SIGMOD 2008 CS848 Paper Presentation Building a Database on S3 Brantner, Florescu, Graf, Kossmann, Kraska SIGMOD 2008 Presented by David R. Cheriton School of Computer Science University of Waterloo 15 March 2010

More information

SCALABLE CONSISTENCY AND TRANSACTION MODELS

SCALABLE CONSISTENCY AND TRANSACTION MODELS Data Management in the Cloud SCALABLE CONSISTENCY AND TRANSACTION MODELS 69 Brewer s Conjecture Three properties that are desirable and expected from realworld shared-data systems C: data consistency A:

More information

Data publication and discovery with Globus

Data publication and discovery with Globus Data publication and discovery with Globus Questions and comments to outreach@globus.org The Globus data publication and discovery services make it easy for institutions and projects to establish collections,

More information

Distributed Systems. Characteristics of Distributed Systems. Lecture Notes 1 Basic Concepts. Operating Systems. Anand Tripathi

Distributed Systems. Characteristics of Distributed Systems. Lecture Notes 1 Basic Concepts. Operating Systems. Anand Tripathi 1 Lecture Notes 1 Basic Concepts Anand Tripathi CSci 8980 Operating Systems Anand Tripathi CSci 8980 1 Distributed Systems A set of computers (hosts or nodes) connected through a communication network.

More information

Distributed Systems. Characteristics of Distributed Systems. Characteristics of Distributed Systems. Goals in Distributed System Designs

Distributed Systems. Characteristics of Distributed Systems. Characteristics of Distributed Systems. Goals in Distributed System Designs 1 Anand Tripathi CSci 8980 Operating Systems Lecture Notes 1 Basic Concepts Distributed Systems A set of computers (hosts or nodes) connected through a communication network. Nodes may have different speeds

More information

Storage and Compute Resource Management via DYRE, 3DcacheGrid, and CompuStore Ioan Raicu, Ian Foster

Storage and Compute Resource Management via DYRE, 3DcacheGrid, and CompuStore Ioan Raicu, Ian Foster Storage and Compute Resource Management via DYRE, 3DcacheGrid, and CompuStore Ioan Raicu, Ian Foster. Overview Both the industry and academia have an increase demand for good policies and mechanisms to

More information

Cloud Computing For Researchers

Cloud Computing For Researchers Cloud Computing For Researchers August, 2016 Compute Canada is often asked about the potential of outsourcing to commercial clouds. This has been investigated as an alternative or supplement to purchasing

More information

Bruce Wright, John Ward, Malcolm Field, Met Office, United Kingdom

Bruce Wright, John Ward, Malcolm Field, Met Office, United Kingdom The Met Office s Logical Store Bruce Wright, John Ward, Malcolm Field, Met Office, United Kingdom Background are the lifeblood of the Met Office. However, over time, the organic, un-governed growth of

More information

The Logical Data Store

The Logical Data Store Tenth ECMWF Workshop on Meteorological Operational Systems 14-18 November 2005, Reading The Logical Data Store Bruce Wright, John Ward & Malcolm Field Crown copyright 2005 Page 1 Contents The presentation

More information

NoSQL systems: sharding, replication and consistency. Riccardo Torlone Università Roma Tre

NoSQL systems: sharding, replication and consistency. Riccardo Torlone Università Roma Tre NoSQL systems: sharding, replication and consistency Riccardo Torlone Università Roma Tre Data distribution NoSQL systems: data distributed over large clusters Aggregate is a natural unit to use for data

More information

Scalable, Reliable Marshalling and Organization of Distributed Large Scale Data Onto Enterprise Storage Environments *

Scalable, Reliable Marshalling and Organization of Distributed Large Scale Data Onto Enterprise Storage Environments * Scalable, Reliable Marshalling and Organization of Distributed Large Scale Data Onto Enterprise Storage Environments * Joesph JaJa joseph@ Mike Smorul toaster@ Fritz McCall fmccall@ Yang Wang wpwy@ Institute

More information

Implementing Trusted Digital Repositories

Implementing Trusted Digital Repositories Implementing Trusted Digital Repositories Reagan W. Moore, Arcot Rajasekar, Richard Marciano San Diego Supercomputer Center 9500 Gilman Drive, La Jolla, CA 92093-0505 {moore, sekar, marciano}@sdsc.edu

More information

Cluster-Based Scalable Network Services

Cluster-Based Scalable Network Services Cluster-Based Scalable Network Services Suhas Uppalapati INFT 803 Oct 05 1999 (Source : Fox, Gribble, Chawathe, and Brewer, SOSP, 1997) Requirements for SNS Incremental scalability and overflow growth

More information

The International Journal of Digital Curation Issue 1, Volume

The International Journal of Digital Curation Issue 1, Volume Towards a Theory of Digital Preservation 63 Towards a Theory of Digital Preservation Reagan Moore, San Diego Supercomputer Center June 2008 Abstract A preservation environment manages communication from

More information

CONSISTENT FOCUS. Building reliable distributed systems at a worldwide scale demands trade-offs between consistency and availability.

CONSISTENT FOCUS. Building reliable distributed systems at a worldwide scale demands trade-offs between consistency and availability. Q SCALABLE FOCUS WEB SERVICES Building reliable distributed systems at a worldwide scale demands trade-offs between consistency and availability. EVENTUALLY CONSISTENT 14 October 2008 ACM QUEUE rants:

More information

Distributed System Chapter 16 Issues in ch 17, ch 18

Distributed System Chapter 16 Issues in ch 17, ch 18 Distributed System Chapter 16 Issues in ch 17, ch 18 1 Chapter 16: Distributed System Structures! Motivation! Types of Network-Based Operating Systems! Network Structure! Network Topology! Communication

More information

Final Exam Review 2. Kathleen Durant CS 3200 Northeastern University Lecture 23

Final Exam Review 2. Kathleen Durant CS 3200 Northeastern University Lecture 23 Final Exam Review 2 Kathleen Durant CS 3200 Northeastern University Lecture 23 QUERY EVALUATION PLAN Representation of a SQL Command SELECT {DISTINCT} FROM {WHERE

More information

L3.4. Data Management Techniques. Frederic Desprez Benjamin Isnard Johan Montagnat

L3.4. Data Management Techniques. Frederic Desprez Benjamin Isnard Johan Montagnat Grid Workflow Efficient Enactment for Data Intensive Applications L3.4 Data Management Techniques Authors : Eddy Caron Frederic Desprez Benjamin Isnard Johan Montagnat Summary : This document presents

More information

Report to Brewer s original presentation of his CAP Theorem at the Symposium on Principles of Distributed Computing (PODC) 2000

Report to Brewer s original presentation of his CAP Theorem at the Symposium on Principles of Distributed Computing (PODC) 2000 Brewer s CAP Theorem Report to Brewer s original presentation of his CAP Theorem at the Symposium on Principles of Distributed Computing (PODC) 2000 Written by Table of Contents Introduction... 2 The CAP-Theorem...

More information

Performance Extrapolation for Load Testing Results of Mixture of Applications

Performance Extrapolation for Load Testing Results of Mixture of Applications Performance Extrapolation for Load Testing Results of Mixture of Applications Subhasri Duttagupta, Manoj Nambiar Tata Innovation Labs, Performance Engineering Research Center Tata Consulting Services Mumbai,

More information

CS555: Distributed Systems [Fall 2017] Dept. Of Computer Science, Colorado State University

CS555: Distributed Systems [Fall 2017] Dept. Of Computer Science, Colorado State University CS 555: DISTRIBUTED SYSTEMS [REPLICATION & CONSISTENCY] Frequently asked questions from the previous class survey Shrideep Pallickara Computer Science Colorado State University L25.1 L25.2 Topics covered

More information

Long-term preservation for INSPIRE: a metadata framework and geo-portal implementation

Long-term preservation for INSPIRE: a metadata framework and geo-portal implementation Long-term preservation for INSPIRE: a metadata framework and geo-portal implementation INSPIRE 2010, KRAKOW Dr. Arif Shaon, Dr. Andrew Woolf (e-science, Science and Technology Facilities Council, UK) 3

More information

Knowledge-based Grids

Knowledge-based Grids Knowledge-based Grids Reagan Moore San Diego Supercomputer Center (http://www.npaci.edu/dice/) Data Intensive Computing Environment Chaitan Baru Walter Crescenzi Amarnath Gupta Bertram Ludaescher Richard

More information

Assignment 5. Georgia Koloniari

Assignment 5. Georgia Koloniari Assignment 5 Georgia Koloniari 2. "Peer-to-Peer Computing" 1. What is the definition of a p2p system given by the authors in sec 1? Compare it with at least one of the definitions surveyed in the last

More information

Experience with Bursty Workflow-driven Workloads in LEAD Science Gateway

Experience with Bursty Workflow-driven Workloads in LEAD Science Gateway Experience with Bursty Workflow-driven Workloads in LEAD Science Gateway Yiming Sun, Suresh Marru, Beth Plale Computer Science Department, Indiana University {yimsun, smarru, plale@cs.indiana.edu ABSTRACT

More information

Outline. Definition of a Distributed System Goals of a Distributed System Types of Distributed Systems

Outline. Definition of a Distributed System Goals of a Distributed System Types of Distributed Systems Distributed Systems Outline Definition of a Distributed System Goals of a Distributed System Types of Distributed Systems What Is A Distributed System? A collection of independent computers that appears

More information

Grid Computing Systems: A Survey and Taxonomy

Grid Computing Systems: A Survey and Taxonomy Grid Computing Systems: A Survey and Taxonomy Material for this lecture from: A Survey and Taxonomy of Resource Management Systems for Grid Computing Systems, K. Krauter, R. Buyya, M. Maheswaran, CS Technical

More information

MOHA: Many-Task Computing Framework on Hadoop

MOHA: Many-Task Computing Framework on Hadoop Apache: Big Data North America 2017 @ Miami MOHA: Many-Task Computing Framework on Hadoop Soonwook Hwang Korea Institute of Science and Technology Information May 18, 2017 Table of Contents Introduction

More information

Oracle Database 10g Resource Manager. An Oracle White Paper October 2005

Oracle Database 10g Resource Manager. An Oracle White Paper October 2005 Oracle Database 10g Resource Manager An Oracle White Paper October 2005 Oracle Database 10g Resource Manager INTRODUCTION... 3 SYSTEM AND RESOURCE MANAGEMENT... 3 ESTABLISHING RESOURCE PLANS AND POLICIES...

More information

A Metadata Catalog Service for Data Intensive Applications

A Metadata Catalog Service for Data Intensive Applications Metadata Catalog Service Draft August 5, 2002 A Metadata Catalog Service for Data Intensive Applications Ann Chervenak, Ewa Deelman, Carl Kesselman, Laura Pearlman, Gurmeet Singh Version 1.0 1 Introduction

More information

Flexible Design for Simple Digital Library Tools and Services

Flexible Design for Simple Digital Library Tools and Services Flexible Design for Simple Digital Library Tools and Services Lighton Phiri Hussein Suleman Digital Libraries Laboratory Department of Computer Science University of Cape Town October 8, 2013 SARU archaeological

More information

BRANCH:IT FINAL YEAR SEVENTH SEM SUBJECT: MOBILE COMPUTING UNIT-IV: MOBILE DATA MANAGEMENT

BRANCH:IT FINAL YEAR SEVENTH SEM SUBJECT: MOBILE COMPUTING UNIT-IV: MOBILE DATA MANAGEMENT - 1 Mobile Data Management: Mobile Transactions - Reporting and Co Transactions Kangaroo Transaction Model - Clustering Model Isolation only transaction 2 Tier Transaction Model Semantic based nomadic

More information

DESIGN AND OVERHEAD ANALYSIS OF WORKFLOWS IN GRID

DESIGN AND OVERHEAD ANALYSIS OF WORKFLOWS IN GRID I J D M S C L Volume 6, o. 1, January-June 2015 DESIG AD OVERHEAD AALYSIS OF WORKFLOWS I GRID S. JAMUA 1, K. REKHA 2, AD R. KAHAVEL 3 ABSRAC Grid workflow execution is approached as a pure best effort

More information

Nowadays data-intensive applications play a

Nowadays data-intensive applications play a Journal of Advances in Computer Engineering and Technology, 3(2) 2017 Data Replication-Based Scheduling in Cloud Computing Environment Bahareh Rahmati 1, Amir Masoud Rahmani 2 Received (2016-02-02) Accepted

More information

Flexible Framework for Mining Meteorological Data

Flexible Framework for Mining Meteorological Data Flexible Framework for Mining Meteorological Data Rahul Ramachandran *, John Rushing, Helen Conover, Sara Graves and Ken Keiser Information Technology and Systems Center University of Alabama in Huntsville

More information

Exploiting peer group concept for adaptive and highly available services

Exploiting peer group concept for adaptive and highly available services Computing in High Energy and Nuclear Physics, 24-28 March 2003 La Jolla California 1 Exploiting peer group concept for adaptive and highly available services Muhammad Asif Jan Centre for European Nuclear

More information

The Use of Cloud Computing Resources in an HPC Environment

The Use of Cloud Computing Resources in an HPC Environment The Use of Cloud Computing Resources in an HPC Environment Bill, Labate, UCLA Office of Information Technology Prakashan Korambath, UCLA Institute for Digital Research & Education Cloud computing becomes

More information

RAID SEMINAR REPORT /09/2004 Asha.P.M NO: 612 S7 ECE

RAID SEMINAR REPORT /09/2004 Asha.P.M NO: 612 S7 ECE RAID SEMINAR REPORT 2004 Submitted on: Submitted by: 24/09/2004 Asha.P.M NO: 612 S7 ECE CONTENTS 1. Introduction 1 2. The array and RAID controller concept 2 2.1. Mirroring 3 2.2. Parity 5 2.3. Error correcting

More information

An Introduction to Big Data Formats

An Introduction to Big Data Formats Introduction to Big Data Formats 1 An Introduction to Big Data Formats Understanding Avro, Parquet, and ORC WHITE PAPER Introduction to Big Data Formats 2 TABLE OF TABLE OF CONTENTS CONTENTS INTRODUCTION

More information

Adapting Mixed Workloads to Meet SLOs in Autonomic DBMSs

Adapting Mixed Workloads to Meet SLOs in Autonomic DBMSs Adapting Mixed Workloads to Meet SLOs in Autonomic DBMSs Baoning Niu, Patrick Martin, Wendy Powley School of Computing, Queen s University Kingston, Ontario, Canada, K7L 3N6 {niu martin wendy}@cs.queensu.ca

More information

Diffusing Your Mobile Apps: Extending In-Network Function Virtualisation to Mobile Function Offloading

Diffusing Your Mobile Apps: Extending In-Network Function Virtualisation to Mobile Function Offloading Diffusing Your Mobile Apps: Extending In-Network Function Virtualisation to Mobile Function Offloading Mario Almeida, Liang Wang*, Jeremy Blackburn, Konstantina Papagiannaki, Jon Crowcroft* Telefonica

More information

Leveraging Software-Defined Storage to Meet Today and Tomorrow s Infrastructure Demands

Leveraging Software-Defined Storage to Meet Today and Tomorrow s Infrastructure Demands Leveraging Software-Defined Storage to Meet Today and Tomorrow s Infrastructure Demands Unleash Your Data Center s Hidden Power September 16, 2014 Molly Rector CMO, EVP Product Management & WW Marketing

More information

FuxiSort. Jiamang Wang, Yongjun Wu, Hua Cai, Zhipeng Tang, Zhiqiang Lv, Bin Lu, Yangyu Tao, Chao Li, Jingren Zhou, Hong Tang Alibaba Group Inc

FuxiSort. Jiamang Wang, Yongjun Wu, Hua Cai, Zhipeng Tang, Zhiqiang Lv, Bin Lu, Yangyu Tao, Chao Li, Jingren Zhou, Hong Tang Alibaba Group Inc Fuxi Jiamang Wang, Yongjun Wu, Hua Cai, Zhipeng Tang, Zhiqiang Lv, Bin Lu, Yangyu Tao, Chao Li, Jingren Zhou, Hong Tang Alibaba Group Inc {jiamang.wang, yongjun.wyj, hua.caihua, zhipeng.tzp, zhiqiang.lv,

More information

dan.fay@microsoft.com Scientific Data Intensive Computing Workshop 2004 Visualizing and Experiencing E 3 Data + Information: Provide a unique experience to reduce time to insight and knowledge through

More information

Performance impact of dynamic parallelism on different clustering algorithms

Performance impact of dynamic parallelism on different clustering algorithms Performance impact of dynamic parallelism on different clustering algorithms Jeffrey DiMarco and Michela Taufer Computer and Information Sciences, University of Delaware E-mail: jdimarco@udel.edu, taufer@udel.edu

More information

THE GLOBUS PROJECT. White Paper. GridFTP. Universal Data Transfer for the Grid

THE GLOBUS PROJECT. White Paper. GridFTP. Universal Data Transfer for the Grid THE GLOBUS PROJECT White Paper GridFTP Universal Data Transfer for the Grid WHITE PAPER GridFTP Universal Data Transfer for the Grid September 5, 2000 Copyright 2000, The University of Chicago and The

More information

Scalable Hybrid Search on Distributed Databases

Scalable Hybrid Search on Distributed Databases Scalable Hybrid Search on Distributed Databases Jungkee Kim 1,2 and Geoffrey Fox 2 1 Department of Computer Science, Florida State University, Tallahassee FL 32306, U.S.A., jungkkim@cs.fsu.edu, 2 Community

More information

SEDA: An Architecture for Well-Conditioned, Scalable Internet Services

SEDA: An Architecture for Well-Conditioned, Scalable Internet Services SEDA: An Architecture for Well-Conditioned, Scalable Internet Services Matt Welsh, David Culler, and Eric Brewer Computer Science Division University of California, Berkeley Operating Systems Principles

More information

QLogic TrueScale InfiniBand and Teraflop Simulations

QLogic TrueScale InfiniBand and Teraflop Simulations WHITE Paper QLogic TrueScale InfiniBand and Teraflop Simulations For ANSYS Mechanical v12 High Performance Interconnect for ANSYS Computer Aided Engineering Solutions Executive Summary Today s challenging

More information

CISC 7610 Lecture 5 Distributed multimedia databases. Topics: Scaling up vs out Replication Partitioning CAP Theorem NoSQL NewSQL

CISC 7610 Lecture 5 Distributed multimedia databases. Topics: Scaling up vs out Replication Partitioning CAP Theorem NoSQL NewSQL CISC 7610 Lecture 5 Distributed multimedia databases Topics: Scaling up vs out Replication Partitioning CAP Theorem NoSQL NewSQL Motivation YouTube receives 400 hours of video per minute That is 200M hours

More information

Dynamically Provisioning Distributed Systems to Meet Target Levels of Performance, Availability, and Data Quality

Dynamically Provisioning Distributed Systems to Meet Target Levels of Performance, Availability, and Data Quality Dynamically Provisioning Distributed Systems to Meet Target Levels of Performance, Availability, and Data Quality Amin Vahdat Department of Computer Science Duke University 1 Introduction Increasingly,

More information

B2SAFE metadata management

B2SAFE metadata management B2SAFE metadata management version 1.2 by Claudio Cacciari, Robert Verkerk, Adil Hasan, Elena Erastova Introduction The B2SAFE service provides a set of functions for long term bit stream data preservation:

More information

Ch 4 : CPU scheduling

Ch 4 : CPU scheduling Ch 4 : CPU scheduling It's the basis of multiprogramming operating systems. By switching the CPU among processes, the operating system can make the computer more productive In a single-processor system,

More information

Multicore Computing and Scientific Discovery

Multicore Computing and Scientific Discovery scientific infrastructure Multicore Computing and Scientific Discovery James Larus Dennis Gannon Microsoft Research In the past half century, parallel computers, parallel computation, and scientific research

More information

VoltDB vs. Redis Benchmark

VoltDB vs. Redis Benchmark Volt vs. Redis Benchmark Motivation and Goals of this Evaluation Compare the performance of several distributed databases that can be used for state storage in some of our applications Low latency is expected

More information

Chapter 10: Performance Patterns

Chapter 10: Performance Patterns Chapter 10: Performance Patterns Patterns A pattern is a common solution to a problem that occurs in many different contexts Patterns capture expert knowledge about best practices in software design in

More information

Chapter 18 Distributed Systems and Web Services

Chapter 18 Distributed Systems and Web Services Chapter 18 Distributed Systems and Web Services Outline 18.1 Introduction 18.2 Distributed File Systems 18.2.1 Distributed File System Concepts 18.2.2 Network File System (NFS) 18.2.3 Andrew File System

More information

Programming model and implementation for processing and. Programs can be automatically parallelized and executed on a large cluster of machines

Programming model and implementation for processing and. Programs can be automatically parallelized and executed on a large cluster of machines A programming model in Cloud: MapReduce Programming model and implementation for processing and generating large data sets Users specify a map function to generate a set of intermediate key/value pairs

More information

Towards Low Overhead Provenance Tracking in Near Real-Time Stream Filtering

Towards Low Overhead Provenance Tracking in Near Real-Time Stream Filtering Towards Low Overhead Provenance Tracking in Near Real-Time Stream Filtering Nithya N. Vijayakumar and Beth Plale Department of Computer Science, Indiana University {nvijayak, plale}@cs.indiana.edu Abstract.

More information

Datacenter replication solution with quasardb

Datacenter replication solution with quasardb Datacenter replication solution with quasardb Technical positioning paper April 2017 Release v1.3 www.quasardb.net Contact: sales@quasardb.net Quasardb A datacenter survival guide quasardb INTRODUCTION

More information

CHAPTER 3 EFFECTIVE ADMISSION CONTROL MECHANISM IN WIRELESS MESH NETWORKS

CHAPTER 3 EFFECTIVE ADMISSION CONTROL MECHANISM IN WIRELESS MESH NETWORKS 28 CHAPTER 3 EFFECTIVE ADMISSION CONTROL MECHANISM IN WIRELESS MESH NETWORKS Introduction Measurement-based scheme, that constantly monitors the network, will incorporate the current network state in the

More information

Consistency and Replication 1/65

Consistency and Replication 1/65 Consistency and Replication 1/65 Replicas and Consistency??? Tatiana Maslany in the show Orphan Black: The story of a group of clones that discover each other and the secret organization Dyad, which was

More information

Lightweight Streaming-based Runtime for Cloud Computing. Shrideep Pallickara. Community Grids Lab, Indiana University

Lightweight Streaming-based Runtime for Cloud Computing. Shrideep Pallickara. Community Grids Lab, Indiana University Lightweight Streaming-based Runtime for Cloud Computing granules Shrideep Pallickara Community Grids Lab, Indiana University A unique confluence of factors have driven the need for cloud computing DEMAND

More information

Building High Performance Apps using NoSQL. Swami Sivasubramanian General Manager, AWS NoSQL

Building High Performance Apps using NoSQL. Swami Sivasubramanian General Manager, AWS NoSQL Building High Performance Apps using NoSQL Swami Sivasubramanian General Manager, AWS NoSQL Building high performance apps There is a lot to building high performance apps Scalability Performance at high

More information

The Computation and Data Needs of Canadian Astronomy

The Computation and Data Needs of Canadian Astronomy Summary The Computation and Data Needs of Canadian Astronomy The Computation and Data Committee In this white paper, we review the role of computing in astronomy and astrophysics and present the Computation

More information

An Integrated Synchronization and Consistency Protocol for the Implementation of a High-Level Parallel Programming Language

An Integrated Synchronization and Consistency Protocol for the Implementation of a High-Level Parallel Programming Language An Integrated Synchronization and Consistency Protocol for the Implementation of a High-Level Parallel Programming Language Martin C. Rinard (martin@cs.ucsb.edu) Department of Computer Science University

More information

Eventual Consistency 1

Eventual Consistency 1 Eventual Consistency 1 Readings Werner Vogels ACM Queue paper http://queue.acm.org/detail.cfm?id=1466448 Dynamo paper http://www.allthingsdistributed.com/files/ amazon-dynamo-sosp2007.pdf Apache Cassandra

More information

Data Management for Distributed Scientific Collaborations Using a Rule Engine

Data Management for Distributed Scientific Collaborations Using a Rule Engine Data Management for Distributed Scientific Collaborations Using a Rule Engine Sara Alspaugh Department of Computer Science University of Virginia alspaugh@virginia.edu Ann Chervenak Information Sciences

More information

Efficient, Scalable, and Provenance-Aware Management of Linked Data

Efficient, Scalable, and Provenance-Aware Management of Linked Data Efficient, Scalable, and Provenance-Aware Management of Linked Data Marcin Wylot 1 Motivation and objectives of the research The proliferation of heterogeneous Linked Data on the Web requires data management

More information

Optimizing the use of the Hard Disk in MapReduce Frameworks for Multi-core Architectures*

Optimizing the use of the Hard Disk in MapReduce Frameworks for Multi-core Architectures* Optimizing the use of the Hard Disk in MapReduce Frameworks for Multi-core Architectures* Tharso Ferreira 1, Antonio Espinosa 1, Juan Carlos Moure 2 and Porfidio Hernández 2 Computer Architecture and Operating

More information

DAMAGE DISCOVERY IN DISTRIBUTED DATABASE SYSTEMS

DAMAGE DISCOVERY IN DISTRIBUTED DATABASE SYSTEMS DAMAGE DISCOVERY IN DISTRIBUTED DATABASE SYSTEMS Yanjun Zuo and Brajendra Panda Abstract Damage assessment and recovery in a distributed database system in a post information attack detection scenario

More information

Evaluation Report: Improving SQL Server Database Performance with Dot Hill AssuredSAN 4824 Flash Upgrades

Evaluation Report: Improving SQL Server Database Performance with Dot Hill AssuredSAN 4824 Flash Upgrades Evaluation Report: Improving SQL Server Database Performance with Dot Hill AssuredSAN 4824 Flash Upgrades Evaluation report prepared under contract with Dot Hill August 2015 Executive Summary Solid state

More information

CA Test Data Manager Key Scenarios

CA Test Data Manager Key Scenarios WHITE PAPER APRIL 2016 CA Test Data Manager Key Scenarios Generate and secure all the data needed for rigorous testing, and provision it to highly distributed teams on demand. Muhammad Arif Application

More information

Joe Wingbermuehle, (A paper written under the guidance of Prof. Raj Jain)

Joe Wingbermuehle, (A paper written under the guidance of Prof. Raj Jain) 1 of 11 5/4/2011 4:49 PM Joe Wingbermuehle, wingbej@wustl.edu (A paper written under the guidance of Prof. Raj Jain) Download The Auto-Pipe system allows one to evaluate various resource mappings and topologies

More information

Users and utilization of CERIT-SC infrastructure

Users and utilization of CERIT-SC infrastructure Users and utilization of CERIT-SC infrastructure Equipment CERIT-SC is an integral part of the national e-infrastructure operated by CESNET, and it leverages many of its services (e.g. management of user

More information

Performance and Scalability with Griddable.io

Performance and Scalability with Griddable.io Performance and Scalability with Griddable.io Executive summary Griddable.io is an industry-leading timeline-consistent synchronized data integration grid across a range of source and target data systems.

More information

3) CHARLIE HULL. Implementing open source search for a major specialist recruiting firm

3) CHARLIE HULL. Implementing open source search for a major specialist recruiting firm Advice: The time spent on pre-launch analysis is worth the effort to avoid starting from scratch and further alienating already frustrated users by implementing a search which appears to have no connection

More information

JULIA ENABLED COMPUTATION OF MOLECULAR LIBRARY COMPLEXITY IN DNA SEQUENCING

JULIA ENABLED COMPUTATION OF MOLECULAR LIBRARY COMPLEXITY IN DNA SEQUENCING JULIA ENABLED COMPUTATION OF MOLECULAR LIBRARY COMPLEXITY IN DNA SEQUENCING Larson Hogstrom, Mukarram Tahir, Andres Hasfura Massachusetts Institute of Technology, Cambridge, Massachusetts, USA 18.337/6.338

More information

Tracking Stream Provenance in Complex Event Processing Systems for Workflow-Driven Computing

Tracking Stream Provenance in Complex Event Processing Systems for Workflow-Driven Computing Tracking Stream Provenance in Complex Event Processing Systems for Workflow-Driven Computing Nithya N. Vijayakumar and Beth Plale Department of Computer Science Indiana University {nvijayak, plale}@cs.indiana.edu

More information

Comparing Open Source Digital Library Software

Comparing Open Source Digital Library Software Comparing Open Source Digital Library Software George Pyrounakis University of Athens, Greece Mara Nikolaidou Harokopio University of Athens, Greece Topic: Digital Libraries: Design and Development, Open

More information

Nuno Freire National Library of Portugal Lisbon, Portugal

Nuno Freire National Library of Portugal Lisbon, Portugal Date submitted: 05/07/2010 UNIMARC in The European Library and related projects Nuno Freire National Library of Portugal Lisbon, Portugal E-mail: nuno.freire@bnportugal.pt Meeting: 148. UNIMARC WORLD LIBRARY

More information

... IBM Advanced Technical Skills IBM Oracle International Competency Center September 2013

... IBM Advanced Technical Skills IBM Oracle International Competency Center September 2013 Performance benefits of IBM Power Systems and IBM FlashSystem for JD Edwards EnterpriseOne IBM Power 780 server with AIX and IBM FlashSystem 820 flash storage improves batch performance in a client proof

More information

Wringing Kilobytes of Knowledge from Petabytes of Data: Something has to Change

Wringing Kilobytes of Knowledge from Petabytes of Data: Something has to Change Wringing Kilobytes of Knowledge from Petabytes of Data: Something has to Change Beth Plale Department of Computer Science School of Informatics Indiana University Research not possible were it not for

More information

MULTIPROCESSORS AND THREAD-LEVEL. B649 Parallel Architectures and Programming

MULTIPROCESSORS AND THREAD-LEVEL. B649 Parallel Architectures and Programming MULTIPROCESSORS AND THREAD-LEVEL PARALLELISM B649 Parallel Architectures and Programming Motivation behind Multiprocessors Limitations of ILP (as already discussed) Growing interest in servers and server-performance

More information

Oracle Database 12c: JMS Sharded Queues

Oracle Database 12c: JMS Sharded Queues Oracle Database 12c: JMS Sharded Queues For high performance, scalable Advanced Queuing ORACLE WHITE PAPER MARCH 2015 Table of Contents Introduction 2 Architecture 3 PERFORMANCE OF AQ-JMS QUEUES 4 PERFORMANCE

More information

ECE 7650 Scalable and Secure Internet Services and Architecture ---- A Systems Perspective

ECE 7650 Scalable and Secure Internet Services and Architecture ---- A Systems Perspective ECE 7650 Scalable and Secure Internet Services and Architecture ---- A Systems Perspective Part II: Data Center Software Architecture: Topic 3: Programming Models Piccolo: Building Fast, Distributed Programs

More information

Optimizing Parallel Access to the BaBar Database System Using CORBA Servers

Optimizing Parallel Access to the BaBar Database System Using CORBA Servers SLAC-PUB-9176 September 2001 Optimizing Parallel Access to the BaBar Database System Using CORBA Servers Jacek Becla 1, Igor Gaponenko 2 1 Stanford Linear Accelerator Center Stanford University, Stanford,

More information

MULTIPROCESSORS AND THREAD-LEVEL PARALLELISM. B649 Parallel Architectures and Programming

MULTIPROCESSORS AND THREAD-LEVEL PARALLELISM. B649 Parallel Architectures and Programming MULTIPROCESSORS AND THREAD-LEVEL PARALLELISM B649 Parallel Architectures and Programming Motivation behind Multiprocessors Limitations of ILP (as already discussed) Growing interest in servers and server-performance

More information

A mining method for tracking changes in temporal association rules from an encoded database

A mining method for tracking changes in temporal association rules from an encoded database A mining method for tracking changes in temporal association rules from an encoded database Chelliah Balasubramanian *, Karuppaswamy Duraiswamy ** K.S.Rangasamy College of Technology, Tiruchengode, Tamil

More information

Hierarchical Chubby: A Scalable, Distributed Locking Service

Hierarchical Chubby: A Scalable, Distributed Locking Service Hierarchical Chubby: A Scalable, Distributed Locking Service Zoë Bohn and Emma Dauterman Abstract We describe a scalable, hierarchical version of Google s locking service, Chubby, designed for use by systems

More information

LAPI on HPS Evaluating Federation

LAPI on HPS Evaluating Federation LAPI on HPS Evaluating Federation Adrian Jackson August 23, 2004 Abstract LAPI is an IBM-specific communication library that performs single-sided operation. This library was well profiled on Phase 1 of

More information