CUBE-wx — User Manual - Forschungszentrum Jülich
Contents
1. CUBE3 User Manual Generic Display for Application Performance Data Version 3 4 March 28 2013 Fengguang Song Felix Wolf Farzona Pulatova Markus Geimer Daniel Becker Brian Wylie Copyright 2008 University of Tennessee Copyright 2008 2013 Forschungszentrum J lich GmbH Contents 1 Introduction 3 2 Using the Display 2 1 Basic Principles 2 2 GUI Components 2 0 wakata aiea ee ee 2 2 1 Tree Browsers Se hg ee i oe He A a ee wl a ee i 2 2 5 Color Legend edo Ae ee ae yh ee e de a ee bid atte ta wanes hae nae pid EE aca en er 2 2 5 Context Menus VD 0 0 0 nt nun A A 2 3 Topology Display 2 3 1 Topology Menu Bar o a 10 Abstract CUBE is a generic presentation component suitable for displaying a wide variety of per formance metrics for parallel programs including MPI and OpenMP applications Program performance is represented in a multi dimensional space including various program and sys tem resources The tool allows the interactive exploration of this space in a scalable fashion and browsing the different kinds of performance behavior with ease CUBE also includes a library to read and write performance data as well as operators to compare integrate and summarize data from different experiments This user manual provides instructions of how to use the CUBE display how to use the operators and how to write CUBE files The CUBE3 implementation has incompatible API and file2. 00 MPI_Attr_get ET 0 00 MPI_Attr_put a ET 0 00 MPI_Bcast 0 00 MPI_Bsend System Tree Topology View Absolute SCs o 0 75 R20 MO Nf J 0 75 R20 M0 Ne Ha 0 75 R20 MO N7 E 0 75 R20 M0 N6 0 62 R20 MO Nd 0 62 R20 MO Nc HI 0 53 R20 MO N5 HI 0 53 R20 MO N4 HI 0 53 R20 MO Nb HI 0 53 R20 M0 Na o 0 62 R20 MO NS E 0 62 R20 M0 N2 0 75 R20 MO N9 E 0 75 R20 M0 N8 E 0 75 R20 MO N 1 A Fa 5 179e 05 0 000 0 0 5 179e 05 Figure 3 CUBE flat profile 2 2 3 Color Legend The color is taken from a spectrum ranging from blue to red representing the whole range of pos sible values To avoid an unnecessary distraction insignificant values close to zero are displayed in dark gray Exact zero values just have the background color 2 2 4 Status Bar The numbers m x n indicate that there are m processes and for each process there are at most n threads in the execution 2 2 5 Context Menus All tree views provide a context menu that can be used to obtain specific information on each node The context menu is accessible via the right mouse button It displays all or a subset of the options described below The call tree has a context menu consisting of two levels The first level menu items are Call site and Called region Choosing the Call site menu shows the information related to the call site and choosing the Called region menu shows the information related t
3. N 1 A ree H 5 179e 05 0 000 0 0 5 179e 05 a 5 179e 05 41 254 124 100 0 41 254 124 8 0 Figure 1 CUBE display window tree each fraction of the severity is displayed only once The purpose of this display strategy is to have a particular performance problem to appear only once in the tree and thus help identify it more quickly Therefore the severity displayed at a node depends on the node s state whether it is expanded or collapsed The severity of a collapsed node represents the whole subtree associated with that node whereas the severity of an expanded node represents only the fraction that is not covered by its descendants because the severity of its descendants is now displayed separately We call the former one inclusive severity whereas we call the latter one exclusive severity H IK 100 main 10 main 30 foo 60 bar Figure 2 Node of the call tree in collapsed or expanded state For instance a call tree may have a node main with two children main foo and main bar Figure 2 In the collapsed state this node is labeled with the time spent in the whole program In the expanded state it displays only the fraction that is spent neither in foo nor in bar Note that the label of a node does not change when it is expanded or collapsed even if the severity of the node changes from exclusive to inclusive or vice versa 2 2 GUI Components The GUI
4. The peer distribution mode shows the percentage relative to the maximum and non zero minimum amount of peer values depending on the current expansion depth The non peer node severity values and all peers with exact zero values are shown as N A Note that in the absolute mode all values are displayed in scientific notation To prevent cluttering the display only the mantissa is shown at the nodes with the exponent displayed at the color legend Each tree view also has a status bar where the left section shows the selected absolute value and the percentage relative to 100 as defined in the selected percentage mode and the right section shows the value or range according to which colors are assigned depending on the selected mode After opening a data set the middle panel shows the call tree of the program However a user might wish to know which fraction of a metric can be attributed to a particular region regardless of from where it was called In this case the user can switch from the call tree mode default to the region profile mode Figure B In the region profile mode the call tree hierarchy is replaced with a source code hierarchy consisting of two levels region and subregions The subregions if applicable are displayed as a single child node labeled subregions A subregions node represents all regions directly called from the region above In this way the user is able to see which fraction of a metric is associated with a region ex
5. consists of a menu bar three tree browsers a color legend and a status bar In addition some tree browsers provides a context menu associated with each node that can be used to access node specific information 2 2 1 Tree Browsers The tree browsers are controlled by the left and right mouse buttons The left mouse button is used to select or expand collapse a node The right mouse button is used to pop up a context menu with node specific information such as online documentation A label in the metric tree shows a metric name A label in the call tree shows the last callee of a particular call path If you want to know the complete call path you must read all labels from the root down to the particular node you are interested in After switching to the region profile view see below labels in the middle tree denote regions depending on their level in the tree A label in the system tree shows the name of the system resource it represents such as a node name or a machine name Processes and threads are usually identified by a number but it is possible to give them specific names when creating a CUBE file The thread level of single threaded applications is hidden Note that all trees can have multiple root nodes Each tree view has its own drop down menu where it is possible to change the way the severty values are displayed The options include absolute value default a root percentage a selection percentage an external percentage a
6. format to preceding versions 1 Introduction CUBE CUBE Uniform Behavioral Encoding is a generic presentation component suitable for displaying a wide variety of performance metrics for parallel programs including MPI and OpenMP 3 applications CUBE allows interactive exploration of a multidimensional metric space in a scalable fashion Scalability is achieved in two ways hierarchical decomposition of individual dimensions and aggregation across different dimensions All metrics are uniformly accommodated in the same display and thus provide the ability to easily compare the effects of different kinds of program behavior CUBE has been designed around a high level data model of program behavior called the CUBE per formance space The CUBE performance space consists of three dimensions a metric dimension a program dimension and a system dimension The metric dimension contains a set of metrics such as communication time or cache misses The program dimension contains the program s call tree which includes all the call paths onto which metric values can be mapped The system dimen sion contains all the control flows of the program which can be processes or threads depending on the parallel programming model Each point m c of the space can be mapped onto a number representing the actual measurement for metric m while the control flow of process thread was executing call path c This mapping is called the severity of the performance s
7. clusively that is without its regions called from there 2 2 2 Menu Bar The menu bar consists of three menus a file menu a view menu and a help menu File The file menu can be used to open and close a file and to exit CUBE It also allows users to add additional mirrors to the existing ones View The view menu can be used to set a reference data set for the external percentage mode If one or more virtual topologies have been defined in the CUBE file and if the user clicks on the topology tab in the GUI the Topology menu item will be enabled Otherwise it is disabled After selecting topolgy tab the Cartesian selection dialog pops up if the CUBE file has multiple topologies Through this dialog users can choose a specific topology view to display in a topology tab next to the system tree tab Please refer to Section 2 3 for detailed information Help Currently the help menu provides only an About dialog with release information O CUBE epik cube File View Help Metrics O 0 00 Time 458519 06 Execution 2 13 MPI 59211 18 Communicat 0 0010 126 38 InitfExit 5 0 00 Overhead E 60000256 Visits A Call Tree Flat Profile ET 0 00 MPI_Abort ET 0 00 MPI_Accumulate ET 0 00 MPI_Address ET 0 00 MPI_Allgather J 0 00 MPI_Allgatherv ET 0 00 MPI_Allreduce 0 00 MPI_Alltoall ET 0 00 MPI_Alltoallv ET 0 00 MPI_Alltoallw ET 0 00 MPI_Attr_delete ET 0
8. cs depend ing on the particular view that has been selected In Figure 1 they represent call paths forming a call tree The nodes in the system dimension represent machines nodes processes or threads from top to bottom Users can perform two types of actions selecting a node or expanding collapsing a node The expansion collapsion behavior for the system tree is different from the other trees because either all entities of a given level are expanded or none Each node is associated with a metric value which is called the severity and is displayed simulta neously using a numerical value as well as a colored square Colors enable the easy identification of nodes of interest even in a large tree whereas the numerical values enable the precise compari son of individual values The sign of a value is visually distinguished by the relief of the colored square A raised relief indicates a positive sign a sunken relief indicates a negative sign A value shown in the metric tree represents the sum of a particular metric for the entire program that is across all call paths and the entire system A value shown in the call tree represents the sum of the selected metric across all processes or threads for a particular call path A value shown in the system tree represents the selected metric for the selected call path and a particular system resource Briefly a tree is always an aggregation of all of its neighbor trees to the right Note that all th
9. e hierarchies in CUBE are inclusion hierarchies meaning that a child node represents a part of the parent node For example the metric hierarchy might display cache misses as a child node of cache accesses because the former event is a subset of the latter event Similarly in Figure 2 the call path main contains the call paths main foo and main bar as child nodes because their execution times are included in their parent s execution time The severity displayed in CUBE follows the principle of single representation that is within a w O 4 CUBE epik cube File View Help Metrics Absoute lt lt C m 0 00 Time Ou 458519 06 Execution a 2 13 MPI L 0 00 Communication W 8374 94 Collective cE 9177 40 P2P 404 72 Late Receiver a O 0 0010 126 38 Init Exit ld 10 66 Synchronization T 0 00 Overhead W 60000256 Visits Call Tree Flat Profile 1 0 00 driver 0 0 00 MPI_Init 0 0 00 MPI_Bcast o0 0 00 MPI_Barrier 0O 0 00 MPI_Cart_create 0 00 inner 0O 0 00 MPI_Barrier E 0 00 sweep T CO 0 00 snd_real 0 00 global_int_sum ET 0 00 MPI_Allreduce ET 0 00 MPI_Finalize System Tree Topology View zami OO BM BG L Lea 2312 75 R20 MO Nf 2403 52 R20 MO Ne 2371 40 R20 MO N7 2462 11 R20 M0 N6 2527 87 R20 M0 Nd 2619 38 R20 M0 Ne Lea 2553 13 R20 MO N5 La 2644 55 R20 M0 N4 2620 59 R20 MO Nb 2711 14 R20 M0 Na 2639 47 R20 M0 N3 2730 50 R20 MO N2 2625 45 R20 MO N9 2717 06 R20 M0 N8 La 2612 00 R20 MO
10. menu and the actual Cartesian grid The Cartesian grid is presented by planes stacked on top of each other in a three dimensional projection The number of planes depends on the number of dimensions in the grid Each plane is divided into squares The number of squares depends on the dimension size Each square represents a system resource e g a process of the application and has a coordinate associate with it The grid displays the severity of the selected metric in the selected call path for each system resource participating in the application s topology The severity is represented as a color A system resource might not be a part of the application s virtual topology or may have a zero value for a metric Therefore it is sometimes possible to have some uncolored squares in the grid picture 2 3 1 Topology Menu Bar The menu related to Topology is located in the View Menu It consists of three submenus a view menu a geometry menu and a zoom menu View The view menu can be used to choose one of the three possible orientations of the grid The coordinate axes at the bottom of the picture indicate the direction of X Y and Z dimensions in the three dimensional space In case of one or two dimensional grids users are provided with only one orientation of the grid Geometry Due to varying dimension sizes planes in the grid might overlap with each other and the size of the squares might be too small to recognize their color This may po
11. n be examined using the CUBE topology view Furthermore the display is augmented with a source code display that can show the exact position of a call site in the source code As performance tuning of parallel applications usually involves multiple experiments to compare the effects of certain optimization strategies CUBE includes a new feature designed to simplify cross experiment analysis The CUBE algebra is an extension of the framework for multi execution performance tuning by Karavanic and Miller and offers a set of operators that can be used to compare integrate and summarize multiple CUBE data sets The algebra allows the combination of multiple CUBE data sets into a single one that can be displayed like the original ones The following sections explain how to use the CUBE display how to create CUBE files and how to use the algebra and other tools 2 Using the Display This section explains how to use the CUBE display component After a brief description of the basic principles different components of the GUI will be described in detail 2 1 Basic Principles The CUBE display consists of three tree browsers each of them representing a dimension of the performance space Figure T The left tree displays the metric dimension the middle tree displays the program dimension and the right tree displays the system dimension The nodes in the metric tree represent metrics The nodes in the program dimension can have different semanti
12. number information to show the correct section of the source code Online description Both metrics and regions can be linked to an online description For example metrics might point to an online documentation explaining their semantics or re gions representing library functions might point to the corresponding library documentation Info A brief description of the selected node supplied by the CUBE data set 2 3 Topology Display In many parallel applications each process or thread communicates only with a limited number of processes The parallel algorithm divides the application domain into smaller chunks known as sub domains A process usually communicates with processes owning sub domains adjacent to its own The mapping of data onto processes and the neighborhood relationship resulting from this mapping is called virtual topology Many applications use one or more virtual topologies Fig ure 4 specified as one two or three dimensional Cartesian grids The CUBE topology display shows performance data mapped onto the Cartesian topology of the application The correspond ing grid is specified by two parameters number of dimensions and size of each dimension v O 4 CUBE e2 cube File View Help Metrics Absolute 12 Time o E 2 00 foo DI 6 00 System time HI 2 00 bar 6 000 25 0 2400e 01 2 000 33 3 Figure 4 Topology Display The display consists of a drop down
13. o the region being called by the call site Oe the callee Location Displays the source code location of a program resource in textual form i e at which line and in what module In the module profile and region profile modes it always refers to the location of its associated region In the call tree mode a call tree node is usually associated with two entities a callsite and the region called by the callsite By entering a specific level of the context menu Callsite or Called region users are able to check either the associated call site s or the called region s location For the call site it shows the call site s location where it has been called or its calling region s location if the line number of the call site is undefined For the called region it shows the location of the region being called by the call site Source code Displays and highlights the source code of a program resource in the source code browser In the module profile and region profile modes it always shows and highlights the source code of its associated region In the call tree mode since each call tree node has a context menu of two levels by choosing the Call site menu it displays and highlights the source code of the call site or the block of source code of the calling region And by choosing the Called region menu it displays and highlights the block of code of the region being called by the call site Note that not all data sets provide sufficient line
14. pace Each dimension of the performance space is organized in a hierarchy First the metric dimension is organized in an inclusion hierarchy where a metric at a lower level is a subset of its parent for example communication time is below execution time Second the program dimension is organized in a call tree hierarchy Flat profiles can be represented as multiple trivial call trees consisting only of a single node Finally the system dimension is organized in a multi level hierarchy consisting of the levels machine SMP node process and thread CUBE also includes a library to read and write instances of the previously described data model in the form of an XML file The file representation is divided into a metadata part and a data part The metadata part describes the structure of the three dimensions plus the definitions of various program and system resources The data part contains the actual severity numbers to be mapped onto the different elements of the performance space The display component can load such a file and display the different dimensions of the perfor mance space using three coupled tree browsers Figure 1 The browsers are connected so that the user can view one dimension with respect to another dimension For example the user can click on a particular metric and see its distribution across the call tree Ifthe CUBE file contains topolog ical information the distribution of the performance metric across the topology ca
15. peer percentage or a peer distribution The last two options are only available for the system tree The absolute value is the real value measured When dis playing a value as a root percentage the percentage refers to the value shown at the root of the metric tree when it is in collapsed state However both absolute mode and root percentage mode have the disadvantage that values can become very small the more you go to the right since aggre gation occurs from right to left To avoid this problem the user can switch to selection percentage Then a percentage in the right or middle tree always refers to the selection in the neighbor to the left that is a percentage in the system dimension refers to the selection in the program dimension and a percentage in the program dimension refers to the selected metric dimension In this mode the percentages in the middle and right tree always sum up to one hundred percent Furthermore to facilitate the comparison of different experiments users can choose the external percentage mode to display percentages relative to another data set The external percentage mode is basically like the normal percentage mode except that the value equal to 100 is determined by another data set The peer percentage mode shows the percentage relative to the maximum amount of peer val ues all entities of the current leaf level depending on the current expansion depth The severity values for the non peer nodes are shown as N A
16. se a problem for the user to view the topology information effectively The geometry menu circumvents this problem by providing options to scale the picture in various ways The Angle option helps the user to adjust the skew of the three dimensional projection The Plane Distance option helps to adjust the inter plane distance The Plane Length option helps users scale the area of each plane Zoom The zoom menu can be used to zoom in or zoom out on the grid References 1 K L Karavanic and B Miller A Framework for Multi Execution Performance Tuning Paral lel and Distributed Computing Practices 4 3 September 2001 Special Issue on Monitoring Systems and Tool Interoperability 2 Message Passing Interface Forum MPI A Message Passing Interface Standard June 1995 http www mpi forum org 3 OpenMP Architecture Review Board OpenMP Application Program Interface Version 2 5 May 2005 http www openmp org 4 E Song F Wolf N Bhatia J Dongarra and S Moore An Algebra for Cross Experiment Performance Analysis In Proc of ICPP 2004 pages 63 72 Montreal Canada August 2004 10
Download Pdf Manuals
Related Search