NOTE. This document is still under development (04/JAN/95). CONTENTS -------- 1. OBTAINING INTERACTIVE GRAPHICAL RESULTS USING THE WORLD WIDE WEB General Description Diagram Showing How to Navigate the GBIS Graphical Interface Pages on the WWW Instructions and Navigational Notes Known bugs / limitations 2. OBTAINING RESULTS BY ANONYMOUS FTP 3. RESULTS FILES NAMING CONVENTION 4. SENDING RESULTS BY ANONYMOUS FTP 5. CONTENTS OF RESULTS FILES 6. CONTENTS OF RESULTS FILES - EXPLANATION OF FORMAT 7. RESULTS FILE STANDARD HEADER 8. TYPE OF RESULTS DATA FOR EACH BENCHMARK 9. SYSTEM DESCRIPTION FOR THE GBIS GRAPHICAL INTERFACE ADMINISTRATOR ________________________________________________________________________________ 1. OBTAINING INTERACTIVE GRAPHICAL RESULTS USING THE WORLD WIDE WEB ------------------------------------------------------------------- (To be completed) General Description ------------------- Diagram Showing How to Navigate the GBIS Graphical Interface Pages on the WWW ----------------------------------------------------------------------------- ___________ |'Home Page'| 1 ___________ | ____________________________________________________ _ _ _ 2 | 3 | 4 | 5 ____________ ______________ _____________________ |'DISCLAIMER'| |'Instructions'| |'Benchmark Selection'| ____________ ______________ _____________________ | 6 | |_____________________________________________| | 7 | ___________________________ | |'List of Available Results'| | 8 ___________________________ ___________________ |'Manufacturer List'| ___________________ | | 9 ___________________ | 'Machine List' | ___________________ | | 10 Do you wish to change the default result format or modify the default graph options :- | ____________________________ | | YES NO | 12 | 11 _________________ | |'CHANGE DEFAULTS'| | _________________ | | | | | | | | | |_____ _____| | | | | | | _____________________________ | RESULT 'Benchmark Name' | 13 _____________________________ NOTES. - Boxes show WWW pages. The document title for these pages is the name shown in quotes preceded by, Southampton HPC Centre - GBIS Graphical Interface - - Most of the pages contain a button labled 'up' to the GBIS Graphical Interface Home Page, which will get you back to the top. In many cases you may wish to go back only one or two pages (e.g. from the final RESULT page to the 'CHANGE DEFAULTS' page, in order to rescale a graph, or from the RESULT page to the 'Machine List' page, in order to re-plot the graph with different machines selected) in which case the navigational buttons of your WWW viewer should be used (e.g. the 'Back' button from mosaic). - Numbers refer to the instructions and navigational notes given below. Instructions and Navigational Notes ----------------------------------- 1. The GBIS Graphical Interface Home Page is accessible on the WWW using the following URL:- http://hpcc.soton.ac.uk/RandD/gbis/papiani-new-gbis-top.html 2. The Home Page contains various links to ... 3. The Home Page contains a link to a Disclaimer page. 4. The Home Page contains a link to this Instructions file (this file is called readme and exists in the pub/benchmark_results directory of the machine containing the Southampton HPC Centre WWW server, refer to section 2. ). 5. The 'Benchmark Selection' page, accessible via a link from the Home Page, is the starting point for selections leading to a graph. This Page is accessible from a link on the Home Page or directly using the following URL:- http://hpcc.soton.ac.uk/RandD/gbis/papiani-new-gbis.html From this page you select a benchmark e.g. Embarrassingly Parallel and then hit the 'Manufacturers' button, which takes you to 8. below. This page also contains 2 links which you can use before selecting the benchmark. These take you to 6. and 7. below (it is useful to access page 7. before selecting a benchmark). 6. The 'Benchmark Selection' page contains a link to the DISCLAIMER page, as in 3. above. 7. The 'Benchmark Selection' page contains a link to the 'List of Available Results' page. This displays as list of all results files. It is useful to scan through the available results before making interactive selections, so that you have some idea of what results are currently available. This screen actually lists all files in subdirectories of the pub/benchmark_results directory (refer to section 2.). These may contain original results files for the benchmark run, not necessarily in the correct format for plotting. Files are listed in alphabetical order of bechmark name, and then in alphabetical order of manufacturers for a particular benchmark. 8. Once a benchmark has been selected in 5. the 'Manufacturer List' page is displayed. This displays a list of manufacturers for which results exist for the chosen benchmark. Select all the manufacturers you are interested in and then hit the 'Machines' button which takes you to 9. below. NOTE. For some benchmarks, this page will also contain additional buttons allowing you to choose the type of results you require. This occurs when a benchmark yields different types of results which can be plotted separately, e.g. for the COMMS1 benchmark, a choice of plots is available; time against message length or performance (bandwidth) against message length. One of the plot types will be selected as a default but you can change this before hitting the 'Machines' button. Different results files exist for different plot types (refer to sections 3. and 8. below), and this selection limits the display of machines (9. below) to those for which the appropriate type of results files exist. 9. After selecting manufacturers of interest in 8. the 'Machine List' page is displayed. This displays the manufacturers selected in 8. as headings. Below each of these headings a list of available machines is displayed. Select the machines of interest and hit the 'Results' button to display the final graph, see 10,11. and 13. below. This page also contains buttons asking if you wish to change default options used for the graph, see 10,12 and 13 below. 10. After selecting the machines of interest in 9. change the buttons on 'Machine List' page from NO to YES, before the hitting the 'Results' button if required. This is set to NO (11.) as a default and must be changed to YES in order to display the 'CHANGE DEFAULTS' page (12. below). 11. Accepting default options in 9. leads directly to the 13. The default options are to autoscale and display the resulting graph as a WWW page (.gif image). 12. The 'CHANGE DEFAULTS' page allows various changes to be made to the results: i) Output format: .gif image - This is the default selection. The graph is returned as a WWW page, as a gif image. .xbm image - The graph is returned as a WWW page as an xbm image (this has a transparent backgroubd). postscript using external viewer - If your viewer is set up to display postscript files using an external viewer (most are) this is the best option to choose. It is quicker and produces a better quality output should you wish to print the graph out. Colour or monochrome postscript can be selected. tabular results - This option does not produce a graph but displays the contents of all the results files for the selected machines. These will contain the xy coordinate values and other comment information. ii) Select ranges for the x and y axes: Enter values for these ranges if you do not wish to accept autoscaling. To enter exponents use e+ or e- as required. iii) Select log or linear for the x and y axes. iv) Move the key to trace names: This appears in the top right-hand corner of the plot by default. Enter xy coordinates if you wish to move the key (e.g. if it is obscuring the traces). The coordinate marks the position of the first symbol in the key. The key can be moved above or below the displayed axes if required. Hit the 'Results' button once you have set the required options. 13. The RESULTS page is displayed, in the default format or according to options set in 12. If an ERROR screen is displayed at any stage use the 'back' button of your viewer to go back and make changes. Improving response time ----------------------- Note. If your WWW viewer can display postscript files using an external viewer then the response time of the GBIS Graphical Interface can be improved by selecting change default options just before you plot a graph, and then selecting the 'postscript using external viewer' format. This is done as follows: i) Once you have selected the machines of interest from the 'Machine List' page, change the buttons labeled 'Do you wish to change the default fromat or modify the default graph options', from NO to YES. ii) Hit the 'Results' button. iii)The 'Change Defaults' page will be displayed. Change the buttons titled 'Select the required output format' so that the 'Display graph as postscript using external viewer' option is selected. Hit the 'Results' button. Known bugs / limitations ------------------------ i) The maximum number of traces on a single plot is limited to around 12 machines. If more machines are selected gnuplot (the graphical package used to generate the graphs) produces an error stating that the maximum input line has been exceeded. Return to the 'Machine List' page and de-select machines as necessary. ________________________________________________________________________________ 2. OBTAINING RESULTS BY ANONYMOUS FTP ------------------------------------- Files sent to the incoming benchmark results directory (see section 3. below) will be renamed and moved to a directory structure where users without access to the GBIS Graphical Interface can obtain results. 1) type ftp par.soton.ac.uk 2) At the 'Name:' prompt, type anonymous 3) Enter your full email address at the password prompt 4) type cd pub/benchmark_results 5) change to sub directories as appropriate - see below 5) type get result_file_name where in step 4) you change directory to a particular benchmark and manufacturer, e.g. /pub/benchmark_results/LPM1/MEIKO this directory will contain a list of files for different MEIKO machines. The filenames will be informative so that users can identify which results they might be interested in, e.g. (Machine Model)_(Date Benchmarked DD.MMM.YY)_(Problem size)_(Benchmark Metric) e.g. CS-2_26.MAY.94_SIZE_ALPHA_2_TEMPO e.g. CS-2_26.MAY.94_SIZE_ALPHA_2_SIMUL (problem size will be included in the name if results files exist for the same machine with the same benchmark date, in order to keep file names unique) See sections 3. and 8. of this document, for further details on file names. ________________________________________________________________________________ 3. RESULTS FILES NAMING CONVENTION ---------------------------------- Results files (in the /pub/benchmark_results directories) have names which end in; _PERF, _TIME, _RATE, _BARTIME, _BARRATE, _AVE, _SEQ_READ, _RAN_READ, _WRITE _BENCH, _TEMPO, _SIMUL, _SPEED, _TRANS, _POLY or _EFFIC These endings identify files which containin xy coordinates. The endings are used to determine axes labels etc. for the particular benchmark - refer to section 8. of this document for further details. File names which do not contain these endings are the original results files for the benchmark run, not necessarily in the correct format for plotting. These are included, when available, as they often contain additional information. Usually the files in graphical format, with the same machine name and date, but with a name ending as above, have been derived from this file. These files can be obtained by anonymous ftp. ________________________________________________________________________________ 4. SENDING RESULTS BY ANONYMOUS FTP ----------------------------------- 1) type ftp par.soton.ac.uk 2) At the 'Name:', prompt type anonymous 3) Enter your full email address at the password prompt 4) type cd incoming/benchmark_results 5) type put result_file_name where result_file_name, would be the file name, containing 5 fields, using the convention:- (Manufacturer)_(machine model)_(date benchmarked DD.MMM.YY)_ (benchmark)_(file number) e.g. MEIKO_CS-2_29.JUL.94_COMMS1_1 file number is incremented for related results files (e.g. results for the same machine, but with different problem sizes or number of processors). Note. Characters within each field should be limited to alphabetic, numeric and - (minus) and . (full-stop) Fields should be separated by _ (underscore) Results files sent in should include the standard header (refer to section 7.) Results will be changed to the correct format (see section 5. below) and renamed (see sections 2. and 3. above) and then moved to the /pub/benchmark_results directories, and will then be accessible by anonymous ftp, and available for use by the GBIS Graphical Interface on the WWW. If the incoming/benchmark_results directory already contains a file with the same name, then increment the file number for the file you wish to send. ________________________________________________________________________________ 5. CONTENTS OF RESULTS FILES ---------------------------- This section gives example graphical results file formats. More detailed information is given in the next section. This section is only relevant if you wish to send results files, in a format that can be immediately incorporated into the results directories. #!, ## and # have a special meaning in results files; explained later. The essential format of a results file is:- #! Machine Model Name ## 2 hashes used for information you want to appear on WWW pages to help ##the user select the required results - this is extracted from each file #single hash interpreted as a comment line 32 2489.8 64 4958.5 128 9980.3 256 19876.2 3 example files follow:- #! T3D ##from NAS Parallel Benchmark report 3-94, received February 94, problem ##size (256**2)*128 #Processors #Performance Mflop/s 32 2489.8 64 4958.5 128 9980.3 256 19876.2 #! Y-MP ##from NAS Parallel Benchmark report 3-94, received August 92, problem ##size (256**2)*128 #Processors #Performance Mflop/s 1 211 8 1677.45 #!C-90 ##from NAS Parallel Benchmark report 3-94, received February 94, problem ##size (256**2)*128 #Processors #Performance Mflop/s 1 1118.3 4 4431 16 17175.4 ________________________________________________________________________________ 6. CONTENTS OF RESULTS FILES - EXPLANATION OF FORMAT ---------------------------------------------------- An example of the file format used by the GBIS Graphical Interface, is given below, the explanation down the right hand side does not form part of the file. Explanation #!CS-2 at the University of Southampton | Text, Machine name | | ##Genesis 3.0, PVM 3.1 Fortran 77, | More text used in ##THU 26 MAY 94, | in providing info ##Compiler version: SC2.0.1 20 Apr | on the choice of ##1993 Sun FORTRAN, | ##2.0.1 patch 100963-03, b/end SC2.0.1 03 Sep | files to plot. ##1992, Operating system: SunOS 5.1 MEIKO | ##FCS | | #Message Length /Byte Transfer Time /us | comment line | 0 1.983E-04 | These lines are 1 2.001E-04 | xy 2 2.006E-04 | coordinates 4 2.007E-04 | #RINF = 9.180 MByte/s | comment lines #NHALF = 1791.949 Byte | containing summary #STARTUP = 195.200 us | parameters There are 5 types of valid lines in a results file. All of these line types will be displayed if text results are chosen (with the initial #, #! or ## characters removed). i) Line beginning with #! This line should contain the machine model name. This is a comment line but is used as the list of machines to choose from i.e. on the 'GBIS Graphical Interface - Machine List' page on the WWW this line appears next to a button so that the machine can be selected. ii) Lines beginning with ## The information in these lines is extracted and displayed in the choice of available files to be plotted, below the machine title (see i) above), i.e. on the 'GBIS Graphical Interface - Machine List' page on the WWW these line appear below the button with the machine name above. These lines should therefore include the benchmark suite, benchmark date and any other information which is important when choosing which files to plot. The Benchmark name and manufacturer do not need to be included in ## lines as this information is displayed as a heading for the group of files. Information in these lines will appear in the same order as it appears in the results files but each line beginning ## will be separated by a space rather than on a new line, when displayed on the WWW. iii)Lines beginning with # These lines are comment lines. It is suggested that the results file header, column headings and any summary parameters are included as comments. These do not appear on the 'Machine List' page on the WWW. These lines do appear if tabular rather than graphical results are selected. iv) Blank lines can be used to improve the layout, but should not be inserted between pairs of xy coordinates. Blank lines should not contain any spaces - only a carriage return. v) Any other lines are interpreted as a pair of xy coordinates. ________________________________________________________________________________ 7. RESULTS FILE STANDARD HEADER ------------------------------- It is suggested that all results files include a standard header as shown below. Fields marked *, should always be completed. For other fields complete as many as possible:- *name: *establishment: *email: telephone: date sent: *date benchmarked (DD/MMM/YY): *benchmark suite and revision: *message passing library and revision: *benchmark: *problem size (if applicable): *number of nodes: *file number if one of a group: *manufacturer: *model number: cpu: cpu speed (MHz): fpu: primary cache: secondary cache: other cache: memory: disk and connection: interconnect type: switching: other hardware: operating system and version: compilers, libraries and versions: compiler switches: various additional tuning parameters: other software: additional comments: where, *file number if one of a group:, indicates the number of related results files (including this one) for the same machine and benchmark, but where the number of processors and/or problem size were changed. e.g. if the Genesis 3.0 FFT1 benchmark is run for 4 problem sizes each run on 1,2,4,16 processors this generates 16 results files. These would eventually be converted to 4 files for the graphical interface - each file would contain Performance versus number of processors, for constant problem size. An example header follows (note lines must begin with # if included as comments in graphical results data files): #*name: Mark Papiani #*establishment: University of Southampton #*email: mp@ecs.soton.ac.uk #telephone: 0703 595000 #date sent: 08/AUG/94 #*date benchmarked (DD/MMM/YY): 15/APR/94 #*benchmark suite and revision: genesis 2.3 #*message passing library and revision: parmacs 5.1 #*benchmark: lpm1 #*problem size (if applicable): alpha=1 #*number of nodes: 8 #*file number if one of a group: 1 of 16 #*manufacturer: MEIKO #*model number: CS-2 #cpu: 8 x scalar I/O processing element (superscalar SPARC) #cpu speed (MHz): 40 MHz #fpu: integrated #primary cache: 20 KByte instruction cache + 16 KByte data cache #secondary cache: none #other cache: none #memory: 32 MBytes #disk and connection: none - swap over internal network to disk. #interconnect type: Meiko - MBUS + custom-chip (ELAN communications processor) #switching: Meiko custom-switch (ELITE) #other hardware: #operating system and version: Solaris 2.3 #compilers, libraries and versions: FORTRAN 2.0.1, Sun Fortran 2.0 libraries #compiler switches: various #additional tuning parameters: #other software: #additional comments: ________________________________________________________________________________ 8. TYPE OF RESULTS DATA FOR EACH BENCHMARK ------------------------------------------ For most Benchmarks Performance / (Mflop/s) is plotted against Processing Elements These results files end in _PERF e.g. CS-2_26.MAY.94_PERF Other file name endings are used where the benchmark uses other axes - see list of benchmarks and available plot types below. The Benchmark Suite for the following Benchmarks is indicated by (G) for a GENESIS Benchmark and (P) for a PARKBENCH Benchmark. LOW LEVEL ----------------- ----------------- TICK1,TICK2 (G,P) ------------------ No graphical information. Text results only. COMMS1,COMMS2,COMMS3 (G,P) ----------------------------------------- For each run two separate graphical results files will exist with the following column headings: #Message Length /Byte Transfer Time /s #Message Length /Byte Transfer Rate /(MByte/s) (file names end in _TIME and _RATE respectively when obtaining results by ftp) Additionally separate results files (with these headings) will be produced if a distinction has been made between long and short messages. RINF1 (G,P) --------- The same type of results as for the COMMS Benchmarks but separate results files for each of the 17 different kernels. filenames contain _KERNEL1 to _KERNEL17 in the name as well as _TIME or _RATE SYNCH1 (G,P) ------------ Results will be grouped into two types of data files: #Time Per Barrier / us Processing Elements #Barrier Rate / (Mbarr/s) Processing Elements (file names end in _BARTIME and _BARRATE respectively when obtaining results by ftp) POLY1,POLY2,POLY3 (G,P) ------------------------------ #RINF / (Mflop/s) Computational Intensity / (flop/mref) (file names end in _POLY when obtaining results by ftp) KERNEL BENCHMARKS ---------------------------------- ---------------------------------- TRANS1 (G) ------ #Performance / (kw_64/s) Processing Elements (file names end in _TRANS when obtaining results by ftp) IO1 (G, possibly P) --- Results will be grouped into four types of data files: Transfer Rate (Write) / (MByte/s) Processing Elements Transfer Rate (Sequential Read) / (MByte/s) Processing Elements Transfer Rate (Random Read) / (MByte/s) Processing Elements Average Transfer Rate / (MByte/s) Processing Elements (file names end in _WRITE , _SEQ_READ , _RAN_READ and _AVE respectively when obtaining results by ftp) FFT1,PDE2,QCD2 (G) ------------------ Files will exist for constant problem size with column headings: #Performance / (Mflop/s) Processing Elements PDE1 (G), SOR (P) ----------------- Files will exist for constant problem size with column headings: #Performance / (Mflop/s) Processing Elements The following PARKBENCH Benchmarks will all use files for constant problem size with column headings: #Performance / (Mflop/s) Processing Elements Dense Matrix Multiply (P) Transpose (P) Dense LU Factorisation with Partial Pivoting (P) QR Decomposition (P) Matrix Tridiagonalisation (P) 3-D FFT (P) multigrid Kernel (P) Embarrassingly Parallel (P) Conjugate Gradient Kernel (P) large integer sort (P) 1-D FFT (P) COMPACT APPLICATIONS -------------------- -------------------- QCD1,GR1 (G) ------------ Files will exist for constant problem size with column headings: #Performance / (Mflop/s) Processing Elements LPM1 (G) ---- Results will be grouped into five types of data files, each for a different performance metric. Each file will contain data for constant problem size: #Temporal Performance / (tstep/s) Processing Elements #Simulation Performance / (sim-ps/s) Processing Elements #Benchmark Performance / (Mflop/s) Processing Elements #Speedup Processing Elements #Efficiency / % Processing Elements (file names end in _TEMPO , _SIMUL , _BENCH , _SPEED and _EFFIC respectively when obtaining results by ftp) MD1 --- Results will be grouped into five types of data files, each for a different performance metric. Each file will contain data for constant problem size: #Temporal Performance / (tstep/s) Processing Elements #Simulation Performance / (sim-ps/s) Processing Elements #Benchmark Performance / (Mflop/s) Processing Elements #Speedup Processing Elements #Efficiency / % Processing Elements (file names end in _TEMPO , _SIMUL and _BENCH respectively when obtaining results by ftp) SOLVER (P) ------ SEIS (P) ---- POLMP (P) ----- Shallow water code (P) ------------------ Molecular dynamics code (P) ----------------------- ... to be completed ____________________________________________________________________________ 9. SYSTEM DESCRIPTION FOR THE GBIS GRAPHICAL INTERFACE ADMINISTRATOR -------------------------------------------------------------------- A description of the system operation is available on par, in directory /pcsg/ftp/pub/benchmark_results/.private Filename:- notes-for-GBIS-Administrator.README -This file is not publicly accessible ____________________________________________________________________________