Name: ravel
Owner: Lawrence Livermore National Laboratory
Description: Ravel MPI trace visualization tool
Created: 2014-12-11 21:45:08.0
Updated: 2018-03-13 19:08:28.0
Pushed: 2018-03-18 19:26:56.0
Size: 1443
Language: C++
GitHub Committers
User | Most Recent Commit | # Commits |
---|
Other Committers
User | Most Recent Commit | # Commits |
---|
Ravel is a trace visualization tool for MPI with recent experimental support for Charm++. Ravel is unique in that it shows not only physical timelines, but also logical ones structured to better capture the intended organization of communication operations. Ravel calculates logical structure from Open Trace Format or Charm++ Projections logs and presents the results using multiple coordinated views.
In logical time, all operations are colored via some metric. The default metric for MPI is lateness which measures the difference in exit time of an operation compared to its peers at the same logical timestep.
Ravel depends on:
To install:
$ git clone https://github.com/scalability-llnl/ravel.git
$ mkdir ravel/build
$ cd ravel/build
$ cmake -DCMAKE_INSTALL_PREFIX=/path/to/install/directory ..
$ make
$ make install
If a dependency is not found, add its install directory to the
CMAKE_PREFIX_PATH
environment variable.
Before opening the trace, check your settings under Options->OTF Importing
.
These options will affect the logical organization even determined by Ravel.
Once you are happy with your options, use File->Open Trace
and navigate to
your .otf
, .otf2
, or .sts
file.
Ravel partitions the trace into fine-grained communication phases – sets of communication operations that must belong together. It imposes a happened-before ordering between traces to better represent how developers think of them separately.
MPI_Isend
s into a single operation
which may send to multiple receive operations. We recommend this option as a
default for all MPI traces.The three timeline views support linked panning, zooming and selection. The overview shows the total metric value over time steps for the whole trace. Clicking and dragging in this view will select a span of timesteps in the other views.
Navigation | Control ———–|——— Pan | Left-click drag Zoom in time | Mouse wheel Zoom in processes | Shift + Mouse wheel Zoom to rectangle | Right-click drag rectangle Select operation | Right-click operation Tool tips | Hover
The cluster view has a slider which changes the size of the neighborhood shown in the upper part of the view. The lower part of the view shows the clusters. Left-click to divide clusters into its children. Click on dendrogram nodes to collapse clusters. Dendrogram pertains to left-most visible partition. Clustering currently shows the first partition rather than all.
All traces are saved in OTF2 and include only the information from the
original trace that is used by Ravel. In addition, communication-related
operations used for logical structure have an OTF2_AttributeList
associated
with their Leave events. These lists include a phase
and step
value
defining the logical structure used by Ravel, as well as any metric values
computed for that operation. Any metric values ending in _agg
represent the
calculated value of the aggregated non-communication operation directly
preceding.
Ravel was written by Kate Isaacs.
Ravel is released under the LGPL license. For more details see the LICENSE file.
LLNL-CODE-663885
Katherine E. Isaacs, Peer-Timo Bremer, Ilir Jusufi, Todd Gamblin, Abhinav Bhatele, Martin Schulz, and Bernd Hamann. Combing the Communication Hairball: Visualizing Parallel Execution Traces using Logical Time. IEEE Transactions on Visualization and Computer Graphics, Proceedings of InfoVis '14, 20(12):2349-2358, December 2014. DOI: 10.1109/TVCG.2014.2346456
Katherine E. Isaacs, Abhinav Bhatele, Jonathan Lifflander, David Boehme, Todd Gamblin, Bernd Hamann, Peer-Timo Bremer. Recovering Logical Structure from Charm++ Event Traces. In Proceedings fo the ACM/IEEE Conference on Supercomputing (SC15), November 2015. DOI: 10.1145/2807591.2807634
Katherine E. Isaacs, Todd Gamblin, Abhinav Bhatele, Martin Schulz, Bernd Hamann, and Peer-Timo Bremer. Ordering Traces Logically to Identify Lateness in Message Passing Programs. IEEE Transactions on Parallel and Distributed Systems, 27(3):829-840, March 2016. DOI: 10.1109/TPDS.2015.2417531