1. A-JUMP Message Passing Framework for Globally Interconnected Clusters Mr. Sajjad Asghar Advance Scientific Computing Group, National Centre for Physics, Islamabad, Pakistan.

2. 6/6/2016CHEP 20102 Research Papers Mehnaz Hafeez, Sajjad Asghar, Usman A. Malik, Adeel-ur-Rehman, and Naveed Riaz, “Message Passing Framework for Globally Interconnected Clusters”, CHEP 2010, Taipei, Taiwan (Accepted). Sajjad Asghar, Mehnaz Hafeez, Usman A. Malik, Adeel-ur-Rehman, and Naveed Riaz, “A Framework for Parallel Code Execution using Java”, IEEE International Conference IEEE International Conference TENCON2010, Fukuoka, Japan (Accepted). Mehnaz Hafeez, Sajjad Asghar, Usman A. Malik, Adeel-ur-Rehman, and Naveed Riaz, “Survey of MPI Implementations delimited by Java”, IEEE International Conference TENCON2010, Fukuoka, Japan (Accepted). Adeel-ur-Rehman, Sajjad Asghar, Usman A. Malik, Mehnaz Hafeez and Naveed Riaz, “Multi-lingual Binding of MPI Code on A-JUMP”, IEEE International Conference on Frontiers of Information Technology, Islamabad Pakistan (Submitted). Sajjad Asghar, Mehnaz Hafeez, Usman A. Malik, Adeel-ur-Rehman, and Naveed Riaz, “A-JUMP, Architecture for Java Universal Message Passing”, IEEE International Conference on Frontiers of Information Technology, Islamabad Pakistan (Submitted).

3. 6/6/2016CHEP 20103 Layout Introduction Related Work A-JUMP Frame work Message Passing for Interconnected Clusters Performance Analysis Future work & Conclusion References

4. 6/6/2016CHEP 20104 Introduction This research work presents a java based message passing framework for globally interconnected clusters.

5. 6/6/2016CHEP 20105 Message Passing Interface Defacto Standard and defines interfaces for C/C++ and Fortran Programmer defines the work and data distribution How and when communication has to be done. Synchronization is implicit (can also be user-defined) Support distributed memory architectures

6. 6/6/2016CHEP 20106 Message Passing Interface Features 1. Standardization 2. Portability 3. Performance 4. Efficient implementations 5. Availability 6. Functionality 7. Scalability The MPI Forum identified some critical shortcomings of existing message passing systems in areas such as complex data layouts or support for modularity and safe communication.

7. 6/6/2016CHEP 20107 Java and MPI Java is an excellent basis for implementing message passing framework. Java supports multi-paradigm communications environment. Java’s bytecode can be executed securely on many platforms. It allows integrating many technologies that are used for existing software development. Java provides many desired features for building wide range of applications like; standalone, client-server, semantic web, enterprise, mobile, graphical or distributed. Thus, it makes Java an excellent basis for achieving productivity of parallel applications.

8. 6/6/2016CHEP 20108 Layout Introduction Related Work A-JUMP Frame work Message Passing for Interconnected Clusters Performance Analysis Future work & Conclusion References

9. 6/6/2016CHEP 20109

10. 6/6/2016CHEP 201010 Architecture for Java Universal Message Passing (A-JUMP) Framework A-JUMP is an HPC framework built for distributed memory architectures in order to facilitate explicit parallelism. A-JUMP is purely written in Java. A-JUMP is based on mpijava1.2 specifications. A-JUMP backbone is called High Performance Computing Bus (HPC Bus).

11. 6/6/2016CHEP 201011 HPC Bus of A-JUMP Communication backbone of the A-JUMP Based on Apache ActiveMQ Includes inter-process communication, monitoring and propagation of Registry information. Provides interoperability between hardware resources and communication protocols. Ensures that any changes in communication protocols and network topologies will remain transparent to end users. Ensures loosely coupled services and components. Segregate the application and communication layers.

12. 6/6/2016CHEP 2010126/6/2016CHEP 201012 HPC Bus of A-JUMP C++ Client

13. 6/6/2016CHEP 201013 Architecture for Java Universal Message Passing (A-JUMP) Framework A-JUMP is designed to provide the following features to write parallel application; Message Passing for Interconnected Clusters Multiple Languages Interoperability Portability Heterogeneity for resource environment Network Topology Independence Multiple Network Protocol Support Communication Mode Autonomy (i.e. asynchronous communication) Different Communication Libraries

14. 6/6/2016CHEP 201014 A-JUMP Component Diagram

15. 6/6/2016CHEP 201015 Layout Introduction Related Work Architecture for Java Universal Message Passing (A-JUMP) Framework Message Passing for Interconnected Clusters Performance Analysis Future work & Conclusion References

16. 6/6/2016CHEP 201016 Grid Computing and MPI GRIDMPI Dynamic Resources can add/delete dynamically Static Static Configuration Host File Voluminous data exchange over Internet Data exchange over cluster Design for internet computingDesign for cluster computing Heterogeneous resource environment Homogenous resource environment

17. 6/6/2016CHEP 201017 Grid Computing and MPI GRIDMPI Specific to certain scientific applications Generic and not specific to certain scientific applications Independent from base Operating System Code execution is Operating System dependent Support for different communication protocols paradigm No support for different communication protocols paradigm High network latencies due to complex layout Low network latencies for message passing

18. 6/6/2016CHEP 201018 A-JUMP, Message Passing for Interconnected Clusters It supports the dynamic behaviour of interconnected clusters with MPI functionality. It is written in pure Java and does not incorporate any native code/libraries. It provides asynchronous mode of communication for both message passing and code distribution over the clusters. It provides sub-clustering of resources for different jobs to utilize free resources efficiently. It is designed for both internet and cluster computing.

19. 6/6/2016CHEP 201019 A-JUMP, Message Passing for Interconnected Clusters

20. HPC Bus ( interconnected clusters) 6/6/2016CHEP 201020 C++ Client

21. 6/6/2016CHEP 201021 A-JUMP, Message Passing for Interconnected Clusters It segregates the communication layer from the application code. It is easy to adopt new communication technologies without changing the application code. It offers support for both homogeneous and heterogeneous clusters for parallel code execution. A set of easy to use APIs are part of the framework.

22. 6/6/2016CHEP 201022 Layered Architecture

23. 6/6/2016CHEP 201023 A-JUMP framework Components Machine Registry Dynamic Resource Discovery Dynamic Resource Allocation Message Security Job Scheduler Monitoring Code Migrator/Execution API Classes

24. 6/6/2016CHEP 201024 Machine Registry Machine Registry provides a list of machines that are dynamically added or removed from the framework. To become the part of the framework, every worker machine has to run Code Migrator/Execution service. Code Migrator/Execution service sends the request to the Machine Registry for a unique rank. Machine Registry assigns unique ranks to the newly added machines. Machine Registry is also responsible to inform Monitoring module about the newly added resources. In future It will records the static information about each machine which has been registered i.e. number of CPUs and number of cores per CPU.

25. 6/6/2016CHEP 201025 Dynamic Resource Discovery (DRD) It keeps track of the resources that are part of the A- JUMP clusters (system size). It gets the information from Machine Registry to index the available machines that are registered with the Machine Registry. DRD fetches the information about the state of the machines from Monitoring. The states are defined as free and busy.

26. 6/6/2016CHEP 201026 Dynamic Resource Allocation (DRA) DRA finds the availability of free resources on A-JUMP clusters. Client sends the request for the required numbers of free processors that are needed to run the job. DRA searches the free available resources from the index maintained by DRD and monitoring. It creates a list of machine that could be used to define a sub cluster from available machines and return this list to client APIs.

27. 6/6/2016CHEP 201027 Message Security It provides the facility to encrypt the outgoing messages. Current implementation is based on simple symmetric encryption technique. Requires more performance analysis

28. 6/6/2016CHEP 201028 Job Scheduler(JS) JS distributes incoming jobs to the cluster. The jobs are sent to the JS by client APIs through ActiveMQ. It works on queuing mechanism which keeps incoming jobs queued until free resources are available for execution. If the number of available resources is less than the required resources then job will not be submitted. Otherwise, a sub-cluster is created by JS according to the requirement. Free machines will remain available for new job request. P2P-MPI is the only other implementation which provides sub- clustering.

29. 6/6/2016CHEP 201029 Monitoring Job Scheduler gets the information from the Monitoring component. Monitoring records the dynamic information about the computational resources that are part of the framework. It records number of busy and free CPUs. It helps the Job Scheduler to find the best match for task in hand. Monitoring maintains two distinct states of the machines available in the cluster; these are free and busy. As soon as a machine starts job execution, it informs Monitoring to change its state from free to busy. Monitoring is also responsible to set the state of the machine back to free as soon as the job gets finish ed.

30. 6/6/2016CHEP 201030 Code Migrator/Execution Code Migrator/Execution distributes and executes a task over the network. A built-in communication API is a part of this component. Enables the tasks to do collective and point to point communication. Both communications between the tasks as well as distribution of code are done in an asynchronous way through ActiveMQ. As soon as a resource gets free, the job is dispatched by Job Scheduler to the designated machine. The underlying code distribution takes place through Code Migrator/Execution. The distribution is done through ActiveMQ. The Code Migrator/Execution on the destination machine accepts the codes and starts its execution. The streamed output is redirected to the user machine. The basic send() and receive() are also incorporated with the help of communication libraries written purely in Java.

31. 6/6/2016CHEP 201031 Code Migrator/Execution Code Adaptor

32. 6/6/2016CHEP 201032 Communication APIs The proposed framework offers easy to use set of communication APIs for writing parallel code. These APIs are close to MPI specifications for Java (mpijava1.2). API Classes Sub cluster Grouping (MPI_COMM_WORLD ) MPI.Init MPI.Finalize MPI.COMM_WORLD.Rank MPI.COMM_WORLD.Size Send Receive Broadcast Gather Scatter Allgather Reduce

33. 6/6/2016CHEP 201033 Workflow of A-JUMP for Interconnected Clusters

34. 6/6/2016CHEP 201034 Figure illustrates A-JUMP relationship and the information shared between the different components. It describes the step by step the flow of program running parallel on different clusters. There are three main entities; Clients Code Adaptors Dynamic Resource Discovery/Dynamic Resource Allocation (DRD/DRA) A client is a process looking for computing resources. A code adaptor is a process offering computing resources Each machine in cluster must run Code Adaptor to receive the incoming program files. The role of a Code Adaptor or a client is not fixed. A machine may serve as a Code Adaptor as well as client. DRD/DRA is a process that allocates the free resources to the client for job submission. DRD/DRA maintains the index of resources and creates sub- clusters based on free computing resources.

35. 6/6/2016CHEP 201035 Workflow of A-JUMP for Multi- Clusters over Internet 1.Registration of New a Machine Each new resource in the cluster is assigned a unique rank by the Machine Registry. Once the new resource is registered, this information is propagated to the Monitoring. 2. Request for Number of processes Client sends a request for number of processes required to run the job. Before the job is submitted to the cluster/clusters, the user sends request to the DRD/DRA for the given number of free processes. 3. Get Free Resources DRD/DRA seeks the information of available free resources from Monitoring. It searches the indexed free processes for sub clustering. 4. Create Sub Cluster Monitoring creates a sub cluster of free available processors and sends it to DRD/DRA. 5. Return Sub Cluster It is the responsibility of DRA/DRD to forward information of the free sub cluster to the client. In case, if required resources are not available then DRD/DRA sends a message about the unavailability of resources to the client. 6. Submit a Job by the Client APIs Once client receives the information about the free required resources, it submits the job to the specific machines identified by the DRD/DRA. These machines may reside on different clusters or on a single cluster. 7. Set Status free or busy When Code Adaptors start execution of the submitted job it informs Machine Registry about the change in the status from free to busy. At the same time Machine Registry also sends the updates in the change of status to the Monitoring. As soon as the submitted job finishes, a request of change in status from busy to free is sent back to machine Registry. Code Adaptors are also responsible to sends results back to the client. 8. Resource Monitoring Information Machine Registry is responsible to send information to Monitoring incase a new resource is added, updated, and deleted. Also when and status gets change from free to busy or vice versa Machine Registry sends the updates to Monitoring.

36. 6/6/2016CHEP 201036 Layout Introduction Related Work Architecture for Java Universal Message Passing (A-JUMP) Framework Message Passing for Interconnected Clusters Performance Analysis Future work & Conclusion References

37. 6/6/2016CHEP 201037 Performance Analysis Communication Performance Measurement Ping Pong Latency Test Network Throughput Measurement Code Speed-up Performance Analysis Embarrassingly Parallel (EP) Benchmark JGF Crypt Benchmark

38. 6/6/2016CHEP 201038 P2P-MPI S. Genaud and C. Rattanapoka, “P2P-MPI: A Peer-to- Peer Framework for Robust Execution of Message Passing Parallel Programs,” Journal of Grid Computing, 5(1):27-42, 2007. Strength Pure Java Supports distributed memory models Supports heterogeneous environment Supports high speed networks Weakness No support for high speed networks Slow Performance

39. 6/6/2016CHEP 201039 Test Environment Results are collected on an environment that simulates globally interconnected clusters. The hardware resources involved in conducting the tests include; 9 single core machines divided into two clusters. The clusters are connected with 1 Gbps bandwidth. The machines are running Windows Server 2003 (with SP2) and Windows XP (with SP3). The TCP window size is default on all the machines in clusters. No optimization is done at the level of OS as well as at the level of hardware.

40. Test Environment 6/6/2016CHEP 201040 Private cluster 1 A-JUMP(HPC Bus) Private cluster 2

41. 6/6/2016CHEP 2010416/6/2016CHEP 201041 Ping Pong Latency Test Time (ms) Data Size (bytes)

42. 6/6/2016CHEP 2010426/6/2016CHEP 201042 Network Throughput Measurement Data Size (bytes) Mbps

43. 6/6/2016CHEP 201043 Code Speed UP Performance Analysis (NAS EP) 6/6/2016CHEP 201043

44. 6/6/2016CHEP 201044 Code Speed UP Performance Analysis (NAS EP)

45. 6/6/2016CHEP 201045 Code Speed UP Performance Analysis (JGF Crypt Class A) 6/6/2016CHEP 201045

46. 6/6/2016CHEP 201046 Code Speed UP Performance Analysis (JGF Crypt Class B) 6/6/2016CHEP 201046

47. 6/6/2016CHEP 201047 Layout Introduction Related Work Architecture for Java Universal Message Passing (A-JUMP) Framework Message Passing for Interconnected Clusters Performance Analysis Conclusion and Future work References

48. 6/6/2016CHEP 201048 Conclusion & Future Work A-JUMP for interconnected clusters is loosely coupled with its communication layer It supports sub clustering It provides dynamic resource management It carries multiple communication protocols It supports asynchronous message passing It supports heterogeneous environment Performance analysis results of A-JUMP For small message sizes are better than P2P-MPI For large message sizes are promising and comparable with P2P-MPI

49. 6/6/2016CHEP 201049 Conclusion & Future Work The communication Library will be improved for message sizes >= 256k Support for different security techniques The other area of focus will be to improve network latencies and lessen the communication overheads Multilingual Support for Interconnected Clusters More NAS and JGF benchmarks

50. 6/6/2016CHEP 201050 Layout Introduction Related Work Architecture for Java Universal Message Passing (A-JUMP) Framework Message Passing for Interconnected Clusters Performance Analysis Conclusion and Future work References

51. 6/6/2016CHEP 201051 References Ian Foster, “The Anatomy of the Grid: Enabling Scalable Virtual Organizations,” in Proc. 7th International Euro-Par Conference on Parallel Processing, Manchester, 2001, pp. 1–4. LCG homepage [Online]. Available: http://lcg.web.cern.ch/lcg. EGEE homepage [Online]. Available: http://www.eu-egee.org. S. Haug, F. Ould-Saada, K. Pajchel, and A. L. Read, “Data management for the world’s largest machine,” in Applied Parallel Computing. State of the Art in Scientific Computing, vol. 4699, Springer Berlin / Heidelberg, 2007, pp. 480-488. P. Eerola et al., “The NorduGrid production grid infrastructure, status and plans,” in IEEE/ACM Grid, Nov. 2003, pp. 158– 165. TeraGrid Project homepage [Online]. Available: http://www.teragrid.org F. Cappello et al., “Grid’5000: a large scale and highly reconfigurable experimental grid testbed.” In Proc. 6th IEEE/ACM International Workhop on Grid Computing - 2005, pp. 99-106. J. Dongarra, D. Gannon, G. Fox, and K. Kennedy. The Impact of Multicore on Computational Science Software. CTWatch Quarterly, 2007 [Online]. Available: http://www.ctwatch.org/quarterly/articles/2007/02/the-impact-of-multicore-on- computational-science-software. S. Genaud and C. Rattanapoka, “P2P-MPI: A Peer-to-Peer Framework for Robust Execution of Message Passing Parallel Programs,” Journal of Grid Computing, 5(1):27-42, 2007. S. Schulz, W. Blochinger, and M. Poths, “A Network Substrate for Peer-to-Peer Grid Computing beyond Embarrassingly Parallel Applications,” in Proc. WRI International Conference on Communications and Mobile Computing – 2009, Volume 03, pp. 60-68. M. Humphrey and M. R. Thompson, “Security Implications of Typical Grid Usage Scenarios,” Cluster Computing, vol. 5, Issue-03, 2002, pp. 257-264. P. Phinjaroenphan, S. Bevinakoppa, and P. Zeephongsekul, “A Method for Estimating the Execution Time of a Parallel Task on a Grid Node,” in Proc. Eurepean Grid Conefernce, Amsterdam, 2005, pp. 226-236. A. Nelisse, J. Maassen, T. Kielmann, and H. Bal. “CCJ: Object-Based Message Passing and Collective Communication in Java,” Concurrency and Computation: Practice & Experience, 15(3-5):341-369, 2003. J. Al-Jaroodi, N. Mohamed, H. Jiang, and D. Swanson, “JOPI: a Java object-passing interface: Research Articles,” Concurrency and Computation: Practice & Experience, vol. 17, Issues 7-8, pp. 775-795, 2005. P. Martin, L. M. Silva, and J. G. Silva, “A Java Interface to WMPI”. In 5th European PVM/MPI Users’ Group Meeting, EuroPVM/MPI’98, Liverpool, UK, Lecture Notes in Computer Science, vol. 1497, pp. 121-128. Springer, 1998. S. Mintchev and V. Getov, “Towards Portable Message Passing in Java: Binding MPI,” in 4th European PVM/MPI Users’ Group Meeting, EuroPVM/MPI’97, Crakow, Poland, Lecture Notes in Computer Science, vol. 1332, pp. 135–142. Springer, 1997. G. Judd, M. Clement, and Q. Snell, “DOGMA: Distributed Object Group Metacomputing Architecture,” Concurrency and Computation: Practice and Experience, 10(11-13):977-983, 1998.

52. 6/6/2016CHEP 201052 References B. Pugh and J. Spacco, “MPJava: High-Performance Message Passing in Java using Java.nio,” in Proc. 16th Intl. Workshop on Languages and Compilers for Parallel Computing (LCPC'03), Lecture Notes in Computer Science, vol. 2958, pp. 323-339, College Station, TX, USA, 2003. B. Y. Zhang, G. W. Yang, and W.-M. Zheng, “Jcluster: an Efficient Java Parallel Environment on a Large-scale Heterogeneous Cluster,” Concurrency and Computation: Practice and Experience, 18(12):1541-1557, 2006. A. Kaminsky, “Parallel Java: A Unified API for Shared Memory and Cluster Parallel Programming in 100% Java” in Proc. 9th Intl. Workshop on Java and Components for Parallelism, Distribution and Concurrency (IWJacPDC'07), pp. 196a (8 pages), Long Beach, CA, USA, 2007. M. Baker, B. Carpenter, G. Fox, S. Ko, and S. Lim, “mpi-Java: an Object-Oriented Java Interface to MPI,” in 1st Int. Workshop on Java for Parallel and Distributed Computing (IPPS/SPDP 1999 Workshop), San Juan, Puerto Rico, Lecture Notes in Computer Science, volume 1586, pp. 748–762. Springer, 1999. S. Genaud and C. Rattanapoka, “P2P-MPI: A Peer-to-Peer Framework for Robust Execution of Message Passing Parallel Programs,” Journal of Grid Computing, 5(1):27-42, 2007. A. Shafi, B. Carpenter, and M. Baker, “Nested Parallelism for Multi-core HPC Systems using Java,” Journal of Parallel and Distributed Computing, 69(6):532-545, 2009. M. Bornemann, R. V. v. Nieuwpoort, and T. Kielmann, “MPJ/Ibis: A Flexible and Efficient Message Passing Platform for Java,” in Proc. 12th European PVM/MPI Users’ Group Meeting (EuroPVM/MPI’05), Lecture Notes in Computer Science, vol. 3666, pp. 217-224, Sorrento, Italy, 2005. S. Bang and J. Ahn, “Implementation and Performance Evaluation of Socket and RMI based Java Message Passing Systems,” in Proc. 5th Intl. Conf. on Software Engineering Research, Management and Applications (SERA'07), pp. 153- 159, Busan, Korea, 2007. G. L. Taboada, J. Tourino, and R. Doallo, “F-MPJ: scalable Java message-passing communications on parallel systems,” Journal of Supercomputing, 2009. V. Getov, G. v. Laszewski, M. Philippsen, and I. Foster, “Multiparadigm communications in Java for grid computing,” Communications of the ACM, Volume. 44, Issue 10, 2001, pp. 118-125. H. Schildt, “Java Beginner’s Guide”, 4th ed., Tata McGraw-Hill, 2007. D. Balkanski, M. Trams, and W. Rehm, “Heterogeneous Computing With MPICH/Madeleine and PACX MPI: a Critical Comparison” Technische Universit, Chemnitz, 2003. S. Asghar, M. Hafeez, U. A. Malik, A. Rehman, and N. Riaz, “A-JUMP, A Universal Parallel Framework,” unpublished. MPI: A Message Passing Interface Standard. Message Passing Interface Forum [Online]. Available: http://www.mpi- forum.org/docs/mpi-11-html/mpi-report.html. G. L. Taboada, J. Tourino, and R. Doallo, “Java for High Performance Computing: Assessment of current research & practice,” in Proc. 7th International Conference on Principles and Practice or Programming in Java, Calgary, Alberta, Canada, 2009. pp. 30-39.

53. 6/6/2016CHEP 201053 References G.L. Taboada, J. Tourino, and R. Doallo, “Performance Analysis of Java Message-Passing Libraries on Fast Ethernet, Myrinet and SCI Clusters,” in Proc. 5th IEEE International Conference on Cluster Computing (Cluster'03), Hong Kong, China, 2003, pp. 118-126. S. Asghar, M. Hafeez, U. A. Malik, A. Rehman, and N. Riaz, “Survey of MPI Implementations delimited by Java for HPC,” unpublished. R. V. v. Nieuwpoort et al., “Ibis: an Efficient Java based Grid Programming Environment,” Concurrency and Computation: Practice and Experience, 17(7-8):1079-1107, 2005. W. Gropp, E. Lusk, N. Doss, and A. Skjellum, “High-performance, portable implementation of the MPI Message Passing Interface Standard. Parallel Computing,” 22(6):789.828, Sept. 1996. MPICH2 [Online]. Available: http://www.mcs.anl.gov/research/projects/mpich2. MPICH-G2 [Online]. Available: http://www3.niu.edu/mpi. N. Karonis, B. Toonen, and I. Foster, “MPICH-G2: A Grid-Enabled Implementation of the Message Passing Interface,” Journal of Parallel and Distributed Computing, 2003. T. Kielmann, R. F. H. Hofman, H. E. Bal, A. Plaat, and R. A. F. Bhoedjang, “MagPIe: MPI's collective communication operations for clustered wide area systems,” ACM SIGPLAN Notices, vol. 34, Issue-08, 1999, pp. 131-140. L. Chen, C. Wang, F. C. M. Lau, and R. K. K. Ma, “A Grid Middleware for Distributed Java Computing with MPI Binding and Process Migration Supports”, Journal of Computer Science and Technology, vol. 18, Issue-04, 2003, pp. 505-514. V. Welch, F. Siebenlist, I. Foster, J. Bresnahan, K. Czajkowski, J. Gawor, C. Kesselman, S. Meder, L. Pearlman, and S. Tuecke, “Security for Grid Services,” Twelfth International Symposium on High Performance Distributed Computing (HPDC- 12), IEEE Press, 2003. Globus Toolkit Homepage [Online]. Available: http://www.globus.org/toolkit. PACX-MPI [Online]. Available: http://www.hlrs.de/organization/av/amt/research/pacx-mpi Version 1.2 of MPI [Online]. Available: http://www.mpi-forum.org/docs/mpi-20-html/node28.htm. O. Aumage, and G. Mercier, "MPICH/Madeleine: a True Multi-Protocol MPI for High Performance Networks," vol. 1, pp.10051a, 15th International Parallel and Distributed Processing Symposium (IPDPS'01), 2001. W. Gropp and E. Lusk, “MPICH working note: The second-generation adi for MPICH implementation of MPI,” 1996. O. Aumage, “Heterogeneous multi-cluster networking with the madeleine III communication library,” in 16th IEEE International Parallel and Distributed Processing Symposium (IPDPS '02 (IPPS & SPDP)), pp 85, Washington - Brussels - Tokyo, Apr 2002. O. Aumage, L. Boug´e, A. Denis, J.-F. M´ehaut, G. Mercier, R. Namyst, and L. Prylli, “A portable and efficient communication library for high-performance cluster computing,” in IEEE Intl Conf. on Cluster Computing (Cluster 2000), pp 78-87, Technische Universitt Chemnitz, Saxony, Germany, Nov 2000. ActiveMQ homepage [Online]. Available: http://activemq.apache.org.

54. 6/6/2016CHEP 201054

55. 6/6/2016CHEP 201055 Class A (3000000 integers) CPUsMessage length (array size) Message SizeNumber of Messages 215000005.72 MB2 47500002.86 MB4 6500 0001.90 MB10 8375 0001.43 MB14

56. 6/6/2016CHEP 201056 Class B (20000000 integers) CPUsMessage length (array size) Message SizeNumber of Messages 210 000 00038.14 MB2 45 000 00019.07 MB4 63 333 33612.71 MB10 82 500 0009.53 MB14