The following repository holds data and source code which is part of the masterproject/thesis of Martin Winter
This code base represents faimGraph: High Performance Management of Fully-dynamic Graphs under tight Memory Constraints on the GPU
IMPORTANT NOTE:
As this is still a research project, the current state is still prone to misconfiguration, hence it is possible to provide the framework with invalid or "bad" values without a warning(e.g. setting the queue size to 0, hence loosing access to all returned indices).
The provided configuration files should provide a guideline on how to set up the project to perform as intended, if questions do arise, it would be highly appreciated to seek contact with the authors for clarification.
The framework currently can handle graphs provided in CSR format as found here: http://www.cc.gatech.edu/dimacs10/
Graphs in other formats would need to be converted to CSR format first.
The framework can be configured using an xml-style configuration file (Description below).
Requirements:
- CUDA 9.1
- gcc-6 or MVSC (VS17) (c++14 support required)
- CMake 3.2 or higher
Linux: To setup and build the project, just run setup.sh
Windows: Use CMake to setup project and build using for example Visual Studio 2017
To run the basic edge update testcase similar to cuSTINGER testcase, run: ./mainaimGraph configuration.xml
Each configuration file contains exactly one global section and at least ONE testrun section, but there can be multiple testruns specified.
global deviceID Which device should be used to run aimGraph (e.g. 0) devicememsize How much memory is allocated by aimGraph (in GB, e.g. 1.75) testrun configuration memorylayout Defines if AOS (Array of Structures) or SOA (Structure of Arrays) should be used Standard version uses AOS graphmode Sets up which graphmode is used simple - weight - semantic updatevariant Can be currently "standard", "warpsized", "vertexcentric" or "vertexcentricsorted" Defines how the update should occur (1 thread or 1 warp per update) or based on affected vertices (+ respecting sort order) rounds Number of rounds for this testrun, the graphstructure will be initialized and torn down this much updaterounds Number of individual update rounds within a round, for each update round batch size of edges is inserted and removed again (can be either the same updates or "realistic" updates, this can be changed in main.cpp via parameter "realisticDeletion") pagesize Size of a page used for edge data initlaunchblocksize Kernel launch size for the Initialization (e.g. 32) insertlaunchblocksize Kernel launch size for the Insertion (e.g. 256) For warpsized approach has to be 32 deletelaunchblocksize Kernel launch size for the Deletion (e.g. 256) For warpsized approach has to be 32 deletionvariation To utilize efficient memory management, "compaction" has to be enabled, only the standard update approach can still perform "standard" deletion (aka no compaction) verification If verification should be performed to check if graph structure on the device matches graph on host (VERY SLOW - unoptimized on CPU) performanceoutput Currently either "csv" or "stdout", changes where the performance output is written to memoryoverallocation Is the overallocation factor that determines how much more memory is allocated per adjacency (should be 1.0, was just introduced for testing purposes) stacksize Stacksize in Bytes queuesize Queuesize in number of items per queue directionality undirected - vs directed Currently not fully implemented graphs filename Can be given multiple times for multiple graphs batches batchsize Can be given multiple times for multiple batchsizes
Performs testruns as:
for (auto batchsize : testrun->batchsizes)
{
for (const auto& graph : testrun->graphs)
{
// Run update procedure on graph with batchsize
for(rounds)
{
// Setup aimGraph
for(updaterounds)
{
// Insert Edgeupdatebatch
// Delete Edgeupdatebatch
}
}
}
}