The Grid:From Parallel to Virtualized Parallel Computing: The Grid: From Parallel to Virtualized Parallel Computing Michael Welzl http://www.welzl.at DPS NSG Team http://dps.uibk.ac.at/nsg Institute of Computer Science University of Innsbruck Habilitation talk TU Darmstadt 14 June 2007


Outline: Outline Grid introduction Middleware first step towards virtualization Research efforts further steps towards virtualization Conclusion


Grid Computing: Grid Computing A brief introduction


Introducing the Grid: Introducing the Grid History: parallel processing at a growing scale Parallel CPU architectures Multiprocessor machines Clusters (“Massively Distributed“) computers on the Internet GRID logical consequence of HPC metaphor: power grid just plug in, don‘t care where (processing) power comes from, don‘t care how it reaches you Common definition: The real and specific problem that underlies the Grid concept is coordinated resource sharing and problem solving in dynamic, multi institutional virtual organizations [Ian Foster, Carl Kesselman and Steven Tuecke, “The Anatomy of the Grid – Enabling Scalable Virtual Organizations”, International Journal on Supercomputer Applications, 2001]


Scope: Scope Definition quite broad (“resource sharing“) Reasonable - e.g., computers also have harddisks But also led to some confusion - e.g., new research areas / buzzwords: Wireless Grid, Data Grid, Semantic / Knowledge Grid, Pervasive Grid, [this space reserved for your favorite research area] Grid Example of confusion due to broad Grid interpretation: “One of the first applications of Grid technologies will be in remote training and education. Imagine the productivity gains if we had routine access to virtual lecture rooms! (..) What if we were able to walk up to a local ‘power wall‘ and give a lecture fully electronically in a virtual environment with interactive Web materials to an audience gathered from around the country - and then simply walk back to the office instead of going back to a hotel or an airplane?“ [I. Foster, C. Kesselman (eds): “The Grid: Blueprint for a New Computing Infrastructure“, 2nd edition, Elsevier Inc. / MKP, 2004]  Clear, narrower scope is advisable for thinking/talking about the Grid Traditional goal: processing power Grid people = parallel people; thus, main goal has not changed much


The next Web?: The next Web? Ways of looking at the Internet Communication medium (email) Truly large kiosk (web) The Grid way of looking at the Internet Infrastructure for Virtual Teams Most of the time... the “real and specific goal“ is High Performance Computing Virtual Organizations and Virtual Teams are well defined i.e. not an „open“ system, e.g. security is a big issue Virtual Teams Geographically distributed Organizationally distributed Yet work on a common problem It has been called“the next web“ But Web 2.0 is already here :-)


Virtual Organizations and Virtual Teams: Virtual Organizations and Virtual Teams Distributed resources and people Linked by networks, crossing admin domains Sharing resources, common goals Dynamic


Austrian Grid E-science Grid applications: Austrian Grid E-science Grid applications Medical Sciences Distributed Heart Simulation Virtual Lung Biopsy Virtual Eye Surgery Medical Multimedia Data Management and Distribution Virtual Arterial Tree Tomography and Morphometry High-Energy Physics CERN experiment analyses Applied Numerical Simulation Distributed Scientific Computing: Advanced Computational Methods in Life Science Computational Engineering High Dimensional Improper Integration Procedures Astrophysical Simulations and Solar Observations Astrophysical Simulations Hydrodynamic Simulations Federation of Distributed Archives of Solar Observation Meteorologal Simulations Environmental GRID Applications


Example: CERN Large Hadron Collider: Example: CERN Large Hadron Collider Largest machine built by humans: particle accelerator and collider with a circumference of 27 kilometers Will generate 10 Petabytes (107 Gigabytes) of information per year … starting 2007! This information must be processed and stored somewhere Beyond the scope of a single institution to manage this problem Projects: LCG (LHC Computing Grid), EGEE (Enabling Grids for E-sciencE)


Complexity: Complexity Grid poses difficult problems Heterogeneity and dynamicity of resources Secure access to resources with different users in various roles, belonging to VTs which belong to VOs Efficient assignment of data and tasks to machines (“scheduling“)


Grid requirements: Grid requirements Computer scientists can tackle these problems Grid application users and programmers are often not computer scientists Important goal: ease of use Programmer should not worry (too much) about the Grid User should worry even less Ultimate goal: write and use an application as if using a single computer (power grid metaphor) How do computer scientists simplify? Abstraction. We build layers. In a Grid, we typically have Middleware.


Grid Middleware: Grid Middleware


Grid computing without middleware: Grid computing without middleware Example manual Grid application execution scp code to 10 machines log in to the 10 machines via ssh and start “application > result“ everywhere Estimate running time, or let application tell you that it‘s done (e.g. via TCP/IP communication in app code) retrieve result files via scp Tedious process - so write a script file Do this again for every application / environment? What if your colleagues need something similar? Standards needed, tools introduced


Toolkits: Toolkits Most famous: Globus Toolkit Evolution from GT2 via GT3 to GT4 influenced the whole Grid community Reference implementation of Open Grid Forum (OGF) standards Other well-known examples Condor Exists since mid-1980‘s No Grid back then - system gradually evolved towards it Traditional goal: harvest CPU power of normal user workstations  many Grid issues always had to be addressed anyway Special interfaces now enable Condor-Globus communication (“Condor-G“) Unicore (used in D-Grid) gLite (used in EGEE) Issues that these middlewares (should) address Load Balancing, error management Authentification, Authorization and Accounting (AAA) Resource discovery, naming Resource access and monitoring Resource reservation and QoS management


Grid Resource Allocation Manager (GRAM): Grid Resource Allocation Manager (GRAM) Globus tool for job execution Unified, resource independent replacement for steps in “manual Grid“ example Unified way to set environment variables: Resource Specification Language (RSL) (stdout = x, arguments = y, ..) Steps 1-4 become Blocking: “globus-job-run -stage hostname applicationname“ -stage option copies code to remote machine Different architectures: recompilation needed – but not supported! Nonblocking: scp code, then “globus-job-submit hostname applicationname“ (staging not yet supported) Obtain unique URL, continuously use it to query job status When done, use “globus-job-get-output URL stdout“ to retrieve stdout More complex systems are built on top of GRAM E.g. Message Passing Interface (MPI) for the Grid: MPICH-G2


GRAM /2: GRAM /2 GRAM leaves a lot of questions unanswered How to recompile application for different architectures? (automatically + in a unified way) What if your computer‘s IP address changes? What if the 10 accessed computer‘s IP addresses change? What if two of the computers becomes unavailable? What if 3 other users start to work with 5 of the 10 computers? A tool for each problem... General-purpose Architecture for Reservation and Allocation (GARA) Integrated QoS via “advance reservation“ of resources (CPU, Disk, Network) Monitoring and Discovery System (MDS) for locating and monitoring resources Resource Broker (Globus: do it yourself; Condor: “matchmaker“) translates requirement specification (CPU, memory, ..) into IP address Diversity of complex tools standardized + available in Globus, addressing some but not all of the issues  need for an architecture


Evolution: moving towards an architecture: Evolution: moving towards an architecture OGSI / OGSA: Open Grid Service Infrastructure / Architecture Open Grid Forum (OGF) standards OGSA = service-oriented architecture; key concept for virtualization use a resource = call a service OGSI = Web Services + state management failed: too complex, not compliant with Web Service standards Source: Globus presentation by Ian Foster


Research towards the power outlet: Research towards the power outlet


Current SoA: Current SoA Standards are only specified when mechanisms are known to work Globus only includes such working elements Lots of important features missing Practical issues with existing middlewares Submitting a Globus job is very slow (Austrian Grid: approx. 20 seconds)  significant granularity limit for parallelization! Globus is a huge piece of software Currently, some confusion about right location of features On top of middleware? (research on top of Globus) In middleware? (other Middleware projects) In the OS? (XtreemOS)  Upcoming slides concern mechanisms which are mostly on top and partially within middleware


Automatic parallelization in Grids: Automatic parallelization in Grids Scheduling; important issue for “power outlet“ goal! Automatic distribution of tasks and inter-task data transmissions = scheduling Grid scheduling encompasses Resource Discovery Authorization Filtering, Application Requirement Definition, Minimal Requirement Filtering System Selection Dynamic Information Gathering System Selection Job Execution (optional) Advance Reservation Job Submission Preparation Tasks Monitoring Progress Job Completion Clean-up Tasks So far, most scheduling efforts consider embarassingly parallel applications - typically parameter sweeps (no dependencies)


Condor case study: Condor case study Application name, parameters, etc. + requirements specified in ClassAds “Requirements = Memory >= 256 && Disk > 10000; Rank = (KFLOPS*10000) + Memory“  only use computers which match requirements (else error), order them by rank Explicit support for parameter sweeps: loop variables Resources registered with description; “central manager“ checks pool against application ClassAds (“matchmaking“) every 5 minutes, assigns jobs Checkpointing in Condor: need to recompile applications, link with special library (redirects syscalls) Save current state for fault tolerance or vacating jobs Because preempted by higher priority job, machine busy, or user demands it Used in Grid Application Development Software Project (GrADS) for rescheduling (dynamic scheduling) and metascheduling (negotiation between multiple applications); ClassAds language extended e.g., aggregation functions such as Max, Min, Sum


Grid workflow applications: Grid workflow applications Dependencies between applications (or large parts of applications) typically specified in Directed Acyclic Graph (DAG) Condor: DAG manager (DAGMan) uses .dag file for simple dependencies “Do not run job ‘B’ until job ‘A’ has completed successfully” DAGMan scheduling: for all tasks do... Find task with earliest starting time Allocate it to processor with Earlierst Finish Time Remove task from list GriPhyN (Grid Physics Network) facilitates workflow design with “Pegasus“ (Planning for Execution in Grids) framework Specification of abstract workflow: identify application components, formulate workflow specifying the execution order, using logical names for components and files Automatic generation of concrete workflow (map components to resources) Concrete workflow submitted to Condor-G/DAGMan


Grid Workflow Applications /2: Grid Workflow Applications /2 Components are built, Web (Grid) Services are defined, Activities are specified Several projects (e.g. K-WF Grid) and systems (e.g. ASKALON) exist Most applications have simple workflows E.g. Montage: dissects space image, distributes processing, merges results


Scheduling example: HEFT algorithmStep 1 - task prioritizing: Scheduling example: HEFT algorithm Step 1 - task prioritizing Rank of a task: longest “distance“ to the end (Mean processing + transfer costs) Tasks are sorted by decreasing rank order


Step 2 - processor selection (EFT): Step 2 - processor selection (EFT) 1 2 4 FT(T1, P1) = 1 FT(T1, P2) = 1 FT(T2, P1) = 1+0.5=1.5 FT(T2, P2) = 1+3+1.5=5.5 FT(T4, P1) = 1.5+1.5=3 FT(T4, P2) = 1.5+2+2.5=6 FT(T3, P1) = 3+2=5 FT(T3, P2) = 1.5+1+2=4.5 FT(T5, P1) = 4.5+2+0.5=7 FT(T5, P2) = 3+7+0.5=10.5 Processor idle + task ready Data transfer Task processing


HEFT discussion: HEFT discussion HEFT is not a solution, just a heuristic problem is known to be NP-complete Outperformed competitors (DAGMan scheduling, genetic algorithm) in ASKALON real-life experiments Still, many improvements possible e.g., other functions than mean, and extension for rescheduling suggested Heterogeneous network capacities and traffic interactions ignored Not detected!


Conclusion: Conclusion


How far have we come?: How far have we come? Remember: systems on last slides are still research Not standardized, not part of reference middleware implementations Right place (OS / Middleware / App) for some functions still undecided A lot is still manual Basically three choices for deploying an application on the Grid Simply use it if it‘s a parameter sweep “Gridify“ it (rewrite using customized allocation - e.g. MPICH-G2) Utilize a workflow tool Convergence between P2P systems and Grids has only just begun Several issues and possible improvements Large number of layers are a mismatch for high performance demands Network usage simplistic, no customized mechanisms


Open issues: layering inefficiencyExample: loss of “connection“ semantics: Open issues: layering inefficiency Example: loss of “connection“ semantics IP TCP HTTP 1.0 SOAP Stateless Connection state Stateless Connection state Web Service Grid Service Doesn‘t care, can do both Stateless Stateful Breaking the chain


Open issues: Open issues Strangely, parallel processing background seems to be ignored E.g., work on task-processor mapping + P2P overlays such as hypercube = ? Microcode Instruction level parallelism Arbitrary parallel applications Parametersweeps Workflow applications


Thank you!: Thank you! Questions?


Backup slides: Backup slides


Research gap: Grid-specificnetwork enhancements: Traditional Internet applications (web browser, ftp, ..) Driving a racing car on a public road Applications with special network properties and requirements Bringing the Grid to its full potential ! Research gap: Grid-specific network enhancements


Grid-network peculiarities: Grid-network peculiarities Special behavior Predictable traffic pattern - this is totally new to the Internet! Web: users create traffic FTP download: starts ... ends Streaming video: either CBR or depends on content! (head movement, ..) Could be exploited by congestion control mechanisms Distinction: Bulk data transfer (e.g. GridFTP) vs. control messages (e.g. SOAP) File transfers are often “pushed“ and not “pulled“ Distributed System which is active for a while overlay based network enhancements possible Multicast P2P paradigm: “do work for others for the sake of enhancing the whole system (in your own interest)“ can be applied - e.g. act as a PEP, ... sophisticated network measurements possible can exploit longevity and distributed infrastructure Special requirements file transfer delay predictions note: useless without knowing about shared bottlenecks QoS, but for file transfers only (“advance reservation“)


What is EC-GIN?: What is EC-GIN? European project: Europe-China Grid InterNetworking STREP in IST FP6 Call 6 2.2 MEuro, 11 partners (7 Europe + 4 China) Networkers developing mechanisms for Grids


Research Challenges: Research Challenges Research Challenges: How to model Grid traffic? Much is known about web traffic (e.g. self-similarity) - but the Grid is different! How to simulate a Grid-network? Necessary for checking various environment conditions May require traffic model (above) Currently, Grid-Sim / Net-Sim are two separate worlds (different goals, assumptions, tools, people) How to specify network requirements? Explicit or implicit, guaranteed or “elastic“, various possible levels of granularity How to align network and Grid economics? Combined usage based pricing for various resources including the network What P2P methods are suitable for the Grid? What is the right means for storing short-lived performance data?


Problem: How Grid people see the Internet: Problem: How Grid people see the Internet Abstraction - simply use what is available still: performance = main goal Wrong. Quote from a paper review: “In fact, any solution that requires changing the TCP/IP protocol stack is practically unapplicable to real-world scenarios, (..).“ How to change this view Create awareness - e.g. GGF GHPN-RG published documents such as “net issues with grids“, “overview of transport protocols“ Develop solutions and publish them! (EC-GIN, GridNets) Just like Web Service community Absolutely not like Web Service community ! Existing transport system (TCP/IP + Routing + ..) works well QoS makes things better, the Grid needs it! we now have a chance for that, thanks to IPv6


A time-to-market issue: A time-to-market issue Result: thesis + running code; tests in collaboration with different research areas Result: thesis + simulation code; perhaps early real-life prototype (if students did well) Typical Grid project Typical Network project


Machine-only communication: Machine-only communication Trend in networks: from support of Human-Human Communication email, chat via Human-Machine Communication web surfing, file downloads (P2P systems), streaming media to Machine-machine Communication Growing number of commercial web service based applications New “hype“ technologies: Sensor nets, Autonomic Computing vision Semantic Web (Services): first big step for supporting machine-only communication at a high level So far, no steps at a lower level This would be like RTP, RTCP, SIP, DCCP, ... for multimedia apps: not absolutely necessary, but advantageous


The long-term value of Grid-net research: The long-term value of Grid-net research A subset of Grid-net developments will be useful for other machine-only communication systems! Key for achieving this: change viewpoint from “what can we do for the Grid“ to “what can the Grid do for us“ (or from “what does the Grid need“ to “what does the Grid mean to us“)


Large stacks: Large stacks IP TCP HTTP SOAP Middleware WS-RF Grid apps


The Grid and P2P systems: The Grid and P2P systems Look quite similar Goal in both cases: resource sharing Major difference: clearly defined VOs / VTs No incentive considerations Availability not such a big problem as in P2P case It is an issue, but at larger time scales (e.g. computers in student labs should be available after 22:00, but are sometimes shut down by tutors) Scalability not such a big issue as in P2P case ...so far!  convergence as Grids grow coordinated resource sharing and problem solving in dynamic, multi institutional virtual organizations (Grid, P2P)


How the tools are applied in practice: How the tools are applied in practice Web Browser Compute Server Data Catalog Data Viewer Tool Certificate authority Chat Tool Credential Repository Web Portal Compute Server Resources implement standard access & management interfaces Collective services aggregate &/or virtualize resources Users work with client applications Application services organize VOs & enable access to other services Database service Database service Database service Simulation Tool Camera Camera Telepresence Monitor Registration Service Source: Globus presentation by Ian Foster


Slide44: Data Mgmt Security Common Runtime Execution Mgmt Info Services Web Services Components Non-WS Components Pre-WS Authentication Authorization GridFTP Pre-WS Grid Resource Alloc. & Mgmt Pre-WS Monitoring & Discovery C Common Libraries Authentication Authorization Reliable File Transfer Data Access & Integration Grid Resource Allocation & Management Index Java WS Core Community Authorization Replica Location eXtensible IO (XIO) Credential Mgmt Community Scheduling Framework Delegation Example: Globus Toolkit version 4 (GT4) Data Replication Trigger C WS Core Python WS Core WebMDS Workspace Management Grid Telecontrol Protocol Contrib/ Preview Core Depre- cated Source: Globus presentation by Ian Foster


Automatic parallelization: Automatic parallelization Has been addressed in the past Microcode parallelism (pipelining in CPU) Relatively easy: simple dependencies Instruction level parallelism More complex dependencies Can automatically be analyzed by compiler Reordering, loop unrolling, .. for (i=1; i<100; i++)   a[i] = a[i] + b[i] * c[i]; /* Thread 1 */ for (i=1; i<50; i++)   a[i] = a[i] + b[i] * c[i];   /* Thread 2 */ for (i=50; i<100; i++)   a[i] = a[i] + b[i] * c[i]; (Intel C++ compiler)


Automatic parallelization /2: Automatic parallelization /2 Parallel Computing: complete applications parallelized Very complex dependencies Decomposition methods + mapping of tasks onto processors: usually not automatic (depends on problem and interconnection network) Algorithm specific methods developed (matrix operations, sorting, ..) Some parts can be automatized, but not everything  explicit parallelism (OpenMP) and even allocation (MPI) quite popular Some research efforts on half-automatic parallelization (“manual“ aid) Programmer knows about problem-specific locality needs (interacting code elements) Examples: Java extensions such as JavaSymphony [Thomas Fahringer, Alexandru Jugravu] HPF+ HALO concept [Siegfried Benkner]


Slide47: Source: http://www.dps.uibk.ac.at/projects/teuta/