A Guide for Proposal Writing Call for Proposals of projects using the HPCI System (General Trial Use project using K computer)

Transcription

1 A Guide for Proposal Writing Call for Proposals of projects using the HPCI System (General Trial Use project using K computer) To apply for the proposal of the projects using the HPCI System, applicants are required to submit the following three documents. (1) Web Application Form (the paper printed out by using the HPCI Online Application System) (2) Application Form (additional sheet 1) [PDF] (3) Program Information (additional sheet 2) [Excel] Please submit a project proposal in accordance with the following procedure. 1. First, please carefully read the Proposal Preparation Instructions. 2. With reference to this Guide for Writing a Proposal from the next page, please fill in the Application Form (additional sheet 1) and the Program Information (additional sheet 2). 3. Using the HPCI Online Application System, please submit the project proposal. Then the applicant needs to upload the Application Form (additional sheet 1) which has been converted to PDF file format and the Program Information (additional sheet 2) of Excel file. For details, please read the latest HPCI Quick Start Guide. 4. After completion of the online submission, please print out the Project application form as the Web Application Form. Then mail the Web Application Form to the mailing address below, with the signature of the Project Representative, the name of the organization to which the Project Representative belongs, and the representative s name of the organization with his/her signature. For mailing, please indicate in red on the envelope that the project proposal is enclosed, and send it by a registered mail. Mailing address RIST Kobe Center Kobe KIMEC Center Building 6F Minatojima-minami-machi, Chuo-ku, Kobe City, Hyogo, Japan Phone: +81-(0) Contact Information (Helpdesk) RIST Kobe Center Minatojima-minami-machi, Chuo-ku, Kobe City, Hyogo, Japan Phone: +81-(0) , FAX +81-(0) helpdesk@hpci-office.jp 1

2 Application Form (additional sheet 1) for General Trial Use for K computer Ver.1.0 Project Name Project Representative Affiliation: Name: 1 Significance and Purpose of the Trial Use The applicant shall briefly indicate an outline of the concept and specific purpose of the research. In particular, describe clearly the following points. 1) Background of the research 2) What will be verified and solved in this trial use? And how large scale computation will be planned in the future research of K computer General Use project? * Please note that preconditions of the General Trial Use project using K computer are: - Using codes which have not run so far in large scale - To test, analyze or tune codes using more than 1000 nodes - For preparation of future application for the General Use project for K computer 2

3 2 Plan and Method of the Trial Use 2.1 Execution Plan and Method of the Trial Use The applicant should provide details of the specific research plan and methods for achieving the Trial Use. Also, for each of the application programs to be used, please describe the reason for selecting the program, the calculation methods (calculation model, analysis method, parallelization techniques, etc.) and the applied verification methods. 2.2 Utilization plan of computational resources Requested computational resources Provider name/resource name: RIKEN/ K computer * Along the research plan, please describe the specific utilization plan, while showing the basis for the estimation (a rough estimate is acceptable) of the computational resource amount to be used. In case you use multiple programs, please clearly indicate the application program that corresponds to the utilization plan. * A maximum resource of 50,000 node hours is available within half a year for the General Trial Use project using K computer. * For the estimation of the requested resource, please do not conflict with the description of the program information (additional sheet 2), for example, regarding i) the correspondence between the maximum value of the number of nodes in the utilization plan and the target value of the number of nodes written in the program information sheet, ii) input data, and iii) the estimated parallel efficiency. < Example of description > An example is shown below, which uses two programs with different calculation methods to verify the calculation method. (Utilization plan) Porting of a program which has shown successful performance by using a PC cluster with Intel CPUs, and the calculation results are verified in small- and middle-class executions. 1. Porting and debugging of the program (Step 1) - the program to be used: ABC_dd - schedule: April - End of May, 201x - expected node-hours (target): 2,000 NH - work to be done: 1) Porting of a program ABC_dd which has showed successful performance by using a PC cluster with Intel CPUs to the K computer. 3

4 2) Performance analysis using a set of verification data, and evaluation of the results. 2. Execution and evaluation using small-scale data (Step 2) - the program to be used: ABC_dd - schedule: June - End of July, 201x - expected node-hours (target): 10,000 NH - work to be done: 1) The program, whose performance will be verified in Step 1 above, will be executed using actual data in small scale. It will be confirmed that the results are the same as those obtained by PC cluster calculations. 2) The program will be tuned, for example, by changing parameters. 3. Test run using middle-scale data (Step 3) - the program to be used: ABC_dd - schedule: August - End of September, 201x - expected node-hours (target): 30,000 NH - work to be done: 1) The program, whose proper performance will be verified using small-scale actual data in Step 2 above, will be executed using actual data in medium scale. It will be confirmed that the results are the same as those obtained by PC cluster calculations. ===================================================== Requested resource amount Provider name / Resource name: RIKEN / the K computer Total : 42,000 node-hours ===================================================== 2.3 Preparation status of the application program Application program name: Regarding the preparation status of the program for carrying out this Trial Use project, please describe the following points for each of the programs to be used. 1) Has the parallelization of the program been finished? Have you verified that the calculation results of the program are correct, by using a parallel computer such as a PC cluster? 2) Have you confirmed that the estimated parallel efficiency for the target value of the number of nodes is sufficient in carrying out the project? (50% or more is expected.) The target value of the number of nodes: Estimated parallel efficiency: % * Please transcribe the target value of the number of nodes and the estimated parallel efficiency from 4

5 the program information (additional sheet 2). * If there is a prospect that enough parallel efficiency will be obtained such as by scale-up of the data, please describe the prospect. 3) If you are going to use a program developed by yourselves, please clearly describe the history and background of the program development. 2.4 Necessity of the requested computational resources * Please describe the necessity of the requested computational resources of the K computer in association with the significance and the research plan of the Trial Use project. 5

6 2.5 Organizational plan for the project execution, and staffing plan Regarding the organization plan for executing the Trial Use project, please describe the following points for each project member: affiliation, specialization, whether or not to use the computing resources, roles and achievements, etc., for each project member. In the staffing plan, please describe the course plan for utilizing workshops, for members who have never use the requested computational resources. Also, please describe the plan if there is a possibility that the member is changed during the project period. * The account of computational resources will be issued to those who have been registered in the WEB site through the HPCI Online Application System. Please match the persons for whom "use" is written in the "Whether to use the resource" field of the Table of project members in this form, with the persons registered in the HPCI Online Application System. <Example of description> (Project Members) Project member Affiliation Specialization Researcher A Researcher B Researcher C External Supporting staff D I University I University I University K Corporation Numerical fluid dynamics, Turbulence Numerical fluid dynamics Computational Science High performance computing Whether to use the resources use use use use Roles Project Representative, Verification of results, Analysis Deputy Project Representative, development of computational models, problem definition, preparation of input data Faster running of the program (Design) Faster tuning of the program (Coding) (Experience and Achievements) Researcher A has past experiences in the field of computational fluid dynamics Researcher B (Staffing plan & training) Apr: B and C will be given the K-computer users Lecture Apr: External support staff D will be contracted. Addition of the project member will be requested. May: D will be given the K-computer users Lecture. 6

7 3 Research information relevant to this proposal * If there is a project that is currently ongoing or was carried out in the past as an HPCI research project, please indicate the "Project ID" and "Project Name", and also describe the status of achievement toward the goal of the project, unresolved issues and possibilities of development in the present project, etc. Project ID: Project Name: Summary: * Research plan for collaborating with large experimental facility projects If the project is collaborating with large experimental facility projects leading to excellent results by complementarily enhancing the effectiveness of the experimental approach with numerical simulations, please mention the collaborative project (including Collaborating Facility Name, Project ID and Project Name, etc.). Collaborating Facility Name: Project ID: Project Name: Summary: 4 Research achievements The applicant should fill in the important research papers/articles, books, industrial property rights, conference presentation and news release that the project members presented up to the present, relating to the research in the sequence of publication from new to old with serial numbers assigned. 1) For published research papers/articles, state the title, the names of authors, name of the journal, referee reading (or not), volume, page numbers of the initial and final pages and the year of publication. 2) The above-mentioned sequence of such items is not compulsory as long as all of the items are included. If there are many authors, only include several authors omitting others (if any authors names are omitted, state the number of the authors and the order on the author list for the related researchers). 7

8 The Project Representative should be marked with a double underline, and the Project member with a single underline. [1] I. Kei, J. Kobe, M. Minatojima, A numerical method for ~, International Journal for Numerical Methods ~, Vol. 78, pp , (2015).( referee reading) [2] News release: World s First Successful turbulence simulation of ~, (2015). [3] I. Kei, J. Kobe, Turbulence analysis of ~, **th ~ Fluid Symposium, Oral presentation, (2014). [4] I. Kei, J. Kobe, Computational Analysis ~, Proceedings of ~, (2013).(without referee reading) 5 Validity to perform the project * Please show that the execution of the project and utilization of the results will be limited to peaceful purposes and proper in terms of Basic Act on Science and Technology and social standards. 8

9 Guide for Writing a Program Information (additional sheet 2) Ver Program Information (additional sheet 2) is an Excel sheet form. Please fill in one sheet per program that you will run in the research project, in one Excel file. 2. You are supposed to fill in blue cells by typing in, you can choose from pull down list for yellow cells, and you can choose from pull down list or fill in by typing in the green cell. Light orange cells are automatically calculated in accordance with your input. 3. Report of program readiness is required. You are supposed to provide this information together with program information. 4. In the cell of actual measurement information, please describe the measured performance by using the same kind of input data a) for your planned production run. Performance information measured in the past could be used. 5. You can choose strong scaling or weak scaling for performance measurement of your program b). a) : the same kind of input data is defined as the data used to perform the same routine or function in the program. b) : Strong scaling is defined as how the elapsed time varies with the number of processors for a fixed problem. Weak scaling is defined as how the elapsed time varies with the number of processors for a fixed problem size per processor. ******************************************************* Description of input items (Red characters required input) ******************************************************* *1 Name of execution machine Please fill in the name of a computer used for performance measurements actually. If a computer of the HPCI computational resources is used, please choose it from the pull down list or input it. *2 Parallelization method If you use MPI, OpenMP or Automatic parallelization to parallelize your program, please select use, and if not, please select no use respectively. If the other parallelization method is used, please describe the name of the method. *3 Measuring method Select either Strong scaling or Weak scaling as the measuring method. 9

10 *4 Information of allocation Please fill in (1) with the number of nodes, (2) with the number of MPI processes in each node, and (3) with the number of multi-threads in each MPI process that you are using for your current simulation. *5 Input data when measured (type and size) Please enter the type and size of input data. The type should be indicated by a subject of your simulation; for example, names of materials, structures, globe or regions around Japan. The size can be indicated by the numbers of data related things such as grids, elements or atoms. Please specify the same type of input data for your planned production run. If the size of input data per process is adjusted to be similar to the size of input data for your planned production run per process, accuracy of the performance estimation will be improved. *6 Execution time Please fill in the cell with execution time (elapsed time) of two cases, parallelization degree m and n (>=2m), in seconds. *7 Effective performance If the floating point operation is a major part of the computational processes, please enter Giga Floating Point Operations Per Second (GFLOPS) and obtain the effective performance by dividing the measured floating point operations per second (FLOPS) by the theoretical peak FLOPS of the computer if possible. These values should be based on the total execution, not on the specific subroutines of the program. If the floating point operations are not major part of computations, please enter the million instructions per second (MIPS) or million operations per second (MOPS) instead of both FLOPS and effective performance. Other units of performance indicators can be chosen in accordance with the type of program execution in answering to this question. If proper measurement tools are not available, you can skip these cells. *8 Usage computational resource Select the name of HPCI computer resources which you are planning to use from the pull down list. *9 Input data for target (type and size) Please enter the type and size of input data for your program to run on your requested resource. *10 Information of allocation for target Please fill in (1) with the number of nodes, (2) with the number of processes in each node, and (3) with the number of threads in each process that are needed for your target simulation. 10

11 ****************************************************************************** From here all values will be automatically calculated in accordance with your input. ****************************************************************************** *11 Degree of parallelism The degree of parallelism will be automatically calculated as the product of the values of three variables; (1), (2), and (3) as shown in *4. *12 Effective parallelization ratio α(%) The effective parallelization ratio is calculated differently depending on the case of strong or weak scaling. The effective parallelization ratio of both strong scaling and weak scaling are calculated automatically in this Excel sheet based on the different formulae. Please utilize these values to set the target of the project (cf. *16 Target degree of parallelism ). The details of the equations are as follows; (a) In case of strong scaling The ratio of the elapsed time of the parallelized part to the total elapsed time and that of the serial part to the total are defined as α and (1-α), respectively. This α will be deduced by employing the Amdahl s law. With this evaluation method, the effective parallelization ratio including the overheads due to parallelization can be obtained. Specifically, the effective parallelization ratio can be obtained through the following calculations. Measure elapsed times, Tm and Tn at two points of parallelism, m and n (n>=2m). Here, we define T1 as an elapsed time with the parallelism of 1. Following Amdahl's law, the ratio of speed-up with the parallelism of m and n can be defined as: SS nn = TT 1 TT nn = SS mm = TT 1 TT mm = 1 1 αα + αα nn 1 1 αα + αα mm (1) (2) By eliminating T1 from equations (1) and (2), the effective parallelization ratio of strong scaling α s is expressed as: αα SS TT mm TT nn α = 1 1 nn TT mm 1 1 mm TT nn (3) If you enter values of Tm and Tn in the Excel Sheet for the case of strong scaling, α s will be calculated automatically by using the equation (3) and shown in terms of %. The calculated effective parallelization ratio might be larger than 1 (unity) or negative, if the 11

12 elapsed time is measured with small m. Please choose proper conditions to avoid such situations. (b) In case of weak scaling The elapsed time of the parallelized part is assumed to be constant. In addition, some part of the serial part is also assumed to be constant. The elapsed time of these parts is assumed to be α in total, and the residual serial part whose elapsed time is dependent on the size of data is assumed to increase in proportion to the size of data. From these assumptions, the relation below holds: TT nn TT mm = αα + (1 αα)nn αα + (1 αα)mm (4) The effective parallelization ratio of weak scaling α w can be solved as follows: α WW αα = nntt mm mmtt nn (1 mm)tt nn (1 nn)tt mm (5) In case of weak scaling measurements of Tm and Tn, α w will be automatically calculated with the equation (5). *13 Parallel Efficiency (En) The parallel efficiency En for a given degree of parallelism n indicates how much the program is efficiently accelerated by parallel processing. En is given by the following formulae. Although their derivation processes are different depending on strong and weak scaling, derived formulae are the same. (a) In case of strong scaling In case of strong scaling, En S is calculated automatically with the following formula, and is displayed in percentage in the sheet: EE SS nn = SS nn nn = 1 (1 αα SS (6) )nn + ααss where α is the effective parallelization ratio, given by the formula (3) with the speed-up ratio Sn, given by Amdahl s law. For example, if n=100, and Sn = 20, 50, and 100 times, En S are 20, 50, and 100%, respectively. If α exceeds 1, En S becomes larger than 100% or less than 0%, and as a result, Amdahl s law is no longer applicable. Then, please try to avoid those situations by properly choosing conditions of performance measurement. (b) In case of weak scaling In case of weak scaling, the amount of computation for a given degree of parallelism n is n times larger than the case with n=1. Ideally, the elapsed time should be the same regardless of whether the software is running in serial or parallel processing. In this case, En W is supposed to be 100%. En W is 50% if the elapsed time in parallel processing is twice as long as that of the serial processing, and 20% if 5 times. Such En W is expressed as a simple ratio of the elapsed time in serial processing to that in parallel processing, and is calculated with a simple formula as follows: 12

13 EE WW nn = TT 1 1 = TT nn (1 αα WW (7) )nn + ααww Here, (7) is derived from (4). Although the ways to derive the effective parallelization ratio α are different depending on strong and weak scaling cases, the formulation of the En is the same in the end in both cases. En W is calculated automatically with the formula (6) or (7), and is displayed in percentage. *14 Guideline for degree of parallelism The degree of parallelism n which gives En = 50% is calculated automatically with the following formula (9). EE nn = 1 (1 αα)nn + αα = 1 2 (8) n = 2 αα 1 αα (9) This value is just a guideline. You can set a larger degree of parallelism as a target, especially if you are planning to speed up your program after starting your project, or if the reduction of the elapsed time is more important than improvement of the parallel efficiency. *15 Target degree of parallelism The target degree of parallelism will be automatically calculated as the product of the values of three variables; (1), (2), and (3) as shown in *10 *16 Estimated parallel efficiency (Ex) The estimated parallel efficiency (Ex) is calculated automatically with the formula (6) or (7) by assigning the target degree of parallelism (x) to n, and the result is shown in %. Please use this value to check whether the value of (x) is appropriate or not. 13