IU ORE-Chem Update

Marlon Pierce, Geoffrey FoxIndiana University

IU to lead New US NSF Track 2d $10M Award

See http://www.futuregrid.org for more information.

What We Said We Would Do• Apply data-centric workflow technologies (Dryad)

– Significant effort• Install and run triple store

– Done locally.– Need to do this in Azure.

• Design alternative formats for ORE (JSON, Microformats)– Nothing to report yet

• Design secure services, compositions, mash-ups– OAuth piece done.– Significant effort on social network interfaces– Nothing to report on ORE-chem enabled services yet

• Investigate clouds for ORE-Chem– Infrastructure and runtime– Significant effort on virtual data stores, overheads of virtualization.

Layer Cake of IU Activities

Web 2.0 Research: Security for REST Services

Cloud Computing: Infrastructure and Runtimes

Infrastructure: Windows HPC Testbeds

Hardware Layer

Cloud Infrastructure• Tempest: HP distributed shared memory cluster with

768 processor cores and 1.5 TB total memory capacity. The cluster includes 13.7 TB of local spinning disk.– Tempest can be dynamically reconfigured to act as either a Windows

HPC or Linux cluster.– Smaller versions Madrid and Barcelona

• Other machines:– The IBM iDataPlex system is an IBM e1350 distributed shared memory

cluster with 1024 processor cores and 3 TB total memory capacity. – Cray XT5m distributed shared memory cluster with 672 processor

cores and 1.3 TB total memory capacity. – A shared memory system with at least 480 cores and 640 GB of RAM

will also be installed at IU as part of the FutureGrid award.

Data Cloud Infrastructure

Triple Store: Intellidimension

• This has been installed on IU servers.• We are ready for data.• Efforts to install this on MS Azure

were not successful.–Inadequate documentation earlier in

the year.–We will revisit this.

Open Elastic Block Store• Amazon EBS is a way to mount virtual disks in cloud-

space.– Empty disk space or archived data stores

– ORECHEM enabled data sets, for example.– Clone-able, so keep your own version of community data.

• We are implementing an open version of this.– Contribute to Nimbus, an open-source EC2– But independent of Xen, etc. – Would be interesting to do this for Windows

• Eventual backbone: IU has over a petabyte disk space of lustre file system. – Can be used to load and store VMs.

• X. Gao won best student poster award at TG09.– Paper accepted to E-Science 2009

Block Store Architecture

Volume Server

Volume Delegate

Virtual Machine Manager (Xen Dom 0)

VMM Delegate

VM instance (Xen Dom U)

VBS Web Service

VBS Client

VBDiSCSI

Create Volume,Export Volume,Create Snapshot,Etc.

Import Volume,Attach Device,Detach Device,Etc.

Integration with Cloud Computing SystemsVolume Server

Volume Delegate

Xen Dom 0Xen

Delegate

Xen Dom U

VBS Web Service

VBS Client

VBDiSCSI

Create Volume,Export Volume,Create Snapshot,Etc.

Import Volume,Attach Device,Detach Device,Etc. Nimbus

Workspace Service

VBS_Nimbus Web Service

Attach-volume <volId> <Nimbus Instance Id> <device>

Query for Xen Dom0 Host and DomUId with <Nimbus Instance Id>

Programming Clouds

Multicore and Cloud Technologies to support Data Intensive applications

• Using Dryad (Microsoft) and MPI to study structure of Gene Sequences on Tempest Cluster. We are working on PubChem.

See http://www.infomall.org/salsa for lab projects (X. Qiu).

PubChem dataset consists of binary 166 MACCS keys (fingerprints), which indicate whether a each chemical compound has a special functional molecule or not We have total 26,466,421 chemical compounds. (i.e, the total PubChem dataset has 166 dimensions and 26M records)

Randomly selected 50K chemicals to produce 3D GTM map. GTM is an algorithm to find a lower dimension structure from higher dimensional data (3D in this case).

http://www.youtube.com/watch?v=nylgjKgnSLg

Courtesy of Jong Y. Choi

IU’s ORE-CHEM PipelineHarvest NIH

PubChem for 3D Structures

Convert PubChem XML

to CML

Convert PubChem XML

to CML

Convert CML to Gaussian Input

Submit Jobs to TeraGrid with

Swarm

Convert Gaussian

Output to CML

Convert CML to RDF->ORE-Chem

Insert RDF into RDF Triple Store

Conversions are done with Jumbo/CML tools from Peter Murray Rust’s group at Cambridge. Swarm is a Web service capable of managing 10,000’s of jobs on the TeraGrid. We are developing a Dryad version of the pipeline.

Goal is to create a public, searchable triple store populated with ORE-CHEM data on drug-like molecules.

Iterative MapReduce- Kmeans Clustering and Matrix Multiplication

Iterative MapReduce algorithm for Matrix Multiplication

Kmeans Clustering implemented as an iterative MapReduce application

Overhead of parallel runtimes – Matrix Multiplication

• Compute intensive application O(n^3)

• Higher data transfer requirements O(n^2)

• CGL-MapReduce shows minimal overheads next to MPI

Overhead of parallel runtimes – Kmeans Clustering • O(n) calculations

in each iteration• Small data transfer

requirements O(1)• With large data

sets, CGL-MapReduce shows negligible overheads

• Extremely higher overheads in Hadoop and Dryad

Jaliya Ekanayake {jekanaya@cs.indiana.edu}

• Performance of MPI on virtualized resources– Evaluated using a dedicated private cloud infrastructure– Exactly the same hardware and software configurations in bare-metal and virtual nodes– Applications with different communication: computation ratios– Different virtual machine(VM) allocation strategies {1-VM per node to 8-VMs per node}

High Performance Parallel Computing on Cloud

Performance of Matrix multiplication under different VM configurations

Overhead under different VM configurations for Concurrent Wave

Equation Solver

• O(n^2) communication (n = dimension of a matrix)

• More susceptible to bandwidth than latency• Minimal overheads under virtualized

resources

• O(1) communication (Smaller messages)

• More susceptible to latency• Higher overheads under virtualized

resourcesJaliya Ekanayake

{jekanaya@cs.indiana.edu}

Conclusions: Dryad for Scientific Computing• Investigated several applications with various computation,

communication, and data access requirements• All DryadLINQ applications work, and in many cases perform

better than Hadoop• We can definitely use DryadLINQ (and Hadoop) for

scientific analyses• We did not implement (find)

– Applications that can only be implemented using DryadLINQ but not with typical MapReduce

• Current release of DryadLINQ has some performance limitations

• DryadLINQ hides many aspects of parallel computing from user

• Coding is much simpler in DryadLINQ than Hadoop (provided that the performance issues are fixed)

• Key issue is support of inhomogeneous data

IU’s ORE-CHEM PipelineHarvest NIH

PubChem for 3D Structures

Convert PubChem XML

to CML

Convert PubChem XML

to CML

Convert CML to Gaussian Input

Submit Jobs to TeraGrid with

Swarm

Convert Gaussian

Output to CML

Convert CML to RDF->ORE-Chem

Insert RDF into RDF Triple Store

Conversions are done with Jumbo/CML tools from Peter Murray Rust’s group at Cambridge. Swarm is a Web service capable of managing 10,000’s of jobs on the TeraGrid. We are developing a Dryad version of the pipeline.

Goal is to create a public, searchable triple store populated with ORE-CHEM data on drug-like molecules.

Architecture of Swarm Service

Windows Server Cluster

Windows Server Cluster

Swarm-Grid

Swarm-Dryad

Local RDMBSLocal RDMBS

Swarm-AnalysisStandard Web Service Interface

Large Task Load Optimizer

Swarm-Grid Connector

Swarm-Dryad Connector

Swarm-Hadoop Connector

Cloud Comp. Cluster

Cloud Comp. Cluster

Grid HPC/Condor Cluster

Grid HPC/Condor Cluster

Swarm-Hadoop

Swarm-Grid• Swarm considers

traditional Grid HPC cluster are suitable for the high-throughput jobs.– Parallel jobs (e.g. MPI

jobs)– Long running jobs

• Resource Ranking Manager– Prioritizes the resources

with QBETS, INCA• Fault Manager

– Fatal faults– Recoverable faults

Resource Ranking Manager

Grid HPC/Condor pool Resource Connector

Condor(Grid/Vanilla) with Birdbath

Grid HPC ClustersGrid HPC ClustersGrid HPC ClustersGrid HPC ClustersGrid HPC ClustersGrid HPC ClustersGrid HPC ClustersGrid HPC Clusters

Condor ClusterCondor Cluster

Standard Web Service Interface

Swarm-Grid

QBETS Web

Service

QBETS Web

Service

Local RDMBS

Local RDMBS

MyProxyServer

MyProxyServer

Hosted by TeraGrid Project

Hosted by UCSB

Request Manager

Job Distributor

Job Queue

Data Model Manager Fault Manager

User A’s Job Board

Some Details

• We can submit jobs to 3 different TeraGrid machines– Abe, Mercury, Cobalt (all at NCSA)– IU’s BigRed has some technical problems

• Can do about 100-200 molecules per day in tests.

• Approach is fragile because application/system admins have tendency to change things every few months.– Don’t validate Globus invocations of codes

Dryad Data Partitioning• Two methods:

– Manually place the files in every node or– Write a C# code that uses DryadLINQ partitioning

operators like Hash Partition<T,K> or Range Partition<T,K>

• A partitioned data set consists of 2 types of files:– A metadata file (.pt as extension) containing metadata

that describes the partitions

– Set of partition files, one for each data partition.

\DryadData\UserName\InputData (file path and name)4 (number of partitions depending on number of nodes available)0,2000,NODE01 (Partition files: Partition number, size(in bytes), node name : File path)1,2000,NODE02,NODE03:FilePath2,2000,NODE03,NODE043,2000,NODE04

Programming the Pipeline• IQueryable<T> represents query over the data• Input data is represented by a PartitionedTable<T> object • DryadLINQ programs apply LINQ query operations to

PartitionedTable<T> objects.• LINQ queries on the PartitionTable object are executed on

the Cluster.

• Jobs will be executed on different nodes and the output would be collected in the outputDirectory.

IQueryable<LineRecord> filenames = PartitionedTable.Get<LineRecord> (filepathuri);IQueryable<outputinfo> outputs = filenames.Select(s => XML2CML( execFileName, programSwitches, s.line, outputDirectory));

Security in Web 2.0

OAuth: REST Security• This is actually a Year 2 deliverable but we made

progress in Year 1.• OAuth is essentially security for REST.

– Provide authentication and authorization– Relevant to ORE-CHEM services

• Use REST URL and HTTP method– Resources are identified by URLs– Access privileges are identified by HTTP methods

(GET, POST, PUT, DELETE)• Extend OAuth

– Add finer-grained authorization information in requests and responses.

OAuth Security Status• OAuth *Core* Code provides the fundamental piece of OAuth

specification 1.0. • Includes minimal webapp example

– The sample web apps just support shared secret.– We extend to support PKI– Also fixed some bugs in the code.– To support OAuth extensions, more code is needed in OAuth core.

• For OpenID, we use library OpenID4Java and it seems to offer enough functionalities so far.

• Tutorial given at TeraGrid09• Slides:

http://www.collab-ogce.org/ogce/images/3/39/OAuthOverview-TG09.ppt

• Code: http://ogce.svn.sourceforge.net/viewvc/ogce/incubator/OGCE-OAuth/

Acknowledgments

• Geoffrey Fox• Judy Qiu and SALSA team: data mining

– www.infomall.org/salsa

• Jaliya Ekanayake: Dryad and Cloud performance• Sangmi Pallickara: Swarm service• Xiaoming Gao: Virtual Block Store• Zhenhua Guo: OAuth, OpenID, and Social

Networks

Full Stop

Dryad and DryadLINQDryad is a high-performance, general-purpose distributed computing engine that simplifies the task of implementing distributed applications on clusters of computers running a windows operating system.

DryadLINQ allows us to implement Dryad applications in managed code by using an extended version of the LINQ programming model and API. LINQ was introduced with Microsoft .NET framework version 3.5.

DryadLINQ provider translates the application’s LINQ queries into a Dryad job and runs the job as a distributed application on a windows HPC cluster. Dryad and DryadLINQ An Introduction version 1.0. June 30, 2009. by Microsoft Research

SS SS S...

Application

...Data Partitions

ProcessingCode

Output Data

HPCCluster

S

Basic Execution

Plan

Distributed Job

Dryad and DryadLINQ An Introduction version 1.0. June 30, 2009. by Microsoft Research

Client Workstation Dryad Cluster

...

DryadLINQProvider Head

NodeCompute Nodes

JM V VV

V - VerticesJM - Dryad Job Manager

DryadLINQ Programming Guide version 1.0. June 30, 2009. by Microsoft Research

Client Workstation: runs DryadLINQ application.

DryadLINQ Provider creates a Windows HPC job on the cluster to handle the Dryad processing, receives the results, and returns them to the application.

Job Manager: Windows HPC task that manages execution of associated Dryad job on the cluster.

Head Node: manages the cluster, hosts the Windows HPC Administration Console and Dryad management service.

Vertices: perform the data processing on the compute nodes.

java.exe + jumboconverters.jar

xml -> cmlcml -> gaussian input

Local Machine

DryadLINQ Provider

(LinqToDryad.dll)

Dryad Cluster

Distribute gaussian Input files across the cluster and run gaussian.exe on every file at every node in the cluster.

Distribute all the initial xml files over the cluster

xml to cml conversionStage1

cml to gaussian conversion

Stage2

run gaussian on every fileStage3

Dryad Cluster

Drilling Though Data Clouds

Bare metal (Computer, network, storage)Bare metal (Computer, network, storage)

FutureGrid/VM/Virtual StorageFutureGrid/VM/Virtual Storage

Cloud Technologies(MapReduce, Dryad, Hadoop)Cloud Technologies(MapReduce, Dryad, Hadoop)

Classic HPCMPIClassic HPCMPI

Applications Cheminformatics: Mapping PubChem data into low dimensions to aid drug discovery Biology: Expressed Sequence Tag (EST) sequence assembly (CAP3) Biology: Pairwise Alu sequence alignment (SW) Health: Correlating childhood obesity with environmental factors

Data mining AlgorithmClustering (Pairwise , Vector)MDS, GTM, PCA, CCA

VisualizationPlotViz

Architecture and Performance of Runtime Environments for Data Intensive Scalable Computing

Data/compute intensive applications implemented as MapReduce “filters”

Architecture of CGL-MapReduce

Measured using 32 Compute nodes each with 8 cores and 16 GB of memory

• Compute intensive application

• Embarrassingly parallel operation

• All runtimes performs equally well

Number of Reads processed

High Energy Physics Data Analysis

CAP3 – Gene Assembly Program

• Data intensive application

• MapReduce style parallel operation

• Both runtimes perform comparably well

Jaliya Ekanayake {jekanaya@cs.indiana.edu}

Job Execution Time in Swarm-Dryad with Windows HPC 16 nodes

Job Execution Time in Swarm-Dryad various number of nodes