4 results about "Out degree" patented technology

The invention discloses a BFS-like algorithm-oriented graph processing system caching method, a controller and a system, and belongs to the technical field of graph processing system performance optimizing.The method comprises the steps that based on the relation between cache page access popularity and the number of unaccessed vertexes in a BFS-like algorithm, more accurate popularity level-to-level management is carried out on cache pages; when the cache is fully loaded, pages with low access potential are effectively eliminated, so that the cache hit rate is increased; fragmented effective edge data are extracted from a cache page and are independently stored in a cache of fine-grained management; the scheduling sequence of the vertexes in the task queue is adjusted by obtaining the proportion of the edge lists of the vertexes in the BFS-like algorithm task queue residing in the cache and combining the attributes such as the out-degree of the vertexes, and the data residing in the cache is utilized to the maximum extent. According to the method, the cache hit rate of the graph processing system when the graph processing system processes the BFS-like algorithm task can be effectively improved, so that the access performance of the graph processing system is improved.

The application provides a code analysis path pruning method and device, equipment and a storage medium, wherein the method comprises the following steps: extracting control flow information of to-be-analyzed code and splitting a composite conditional statement, generating a control flow graph corresponding to the to-be-analyzed code, a node block of the control flow graph representing an execution unit or a conditional judgment unit, and an edge representing a path direction of the control flow; marking a conditional judgment unit to be tracked as a record unit; traversing the control flow graph in a multi-branch parallel mode, performing normalization processing on the statement in the traversal process, if the conditional statement of a certain record unit is first appeared in a single branch, maintaining all possible conditional states, otherwise, querying the value of the conditional state corresponding to the conditional statement from a dynamically maintained list, and pruning the out-degree edge of the record unit which does not match the conditional state; for the first appearance of the conditional statement and the unpruned condition after the query, a copy of the dynamically maintained list is generated, which is used for independent update and query in the subsequent conditional state of each branch.

The invention discloses a clustering rearrangement sampling method for accelerating graph neural network training. The method comprises the following steps: storing an original graph edge sequence in a COO format; performing algorithm clustering on nodes in the graph according to given parameters; rearranging the edges according to the obtained clustering result; calculating the available memory of the GPU, and determining the number of cache nodes; loading the data into a GPU (Graphics Processing Unit) memory; and performing GNN subsequent sampling and calculation operations. The invention aims at providing a clustering rearrangement sampling method for accelerating graph neural network training aiming at the current situation that clustering rearrangement preprocessing is not carried out on GNN sampling at present. The core idea of the method is that clustering is carried out before training, high-similarity nodes are divided into the same cluster, and edge sets in a graph are rearranged in the cluster. The rearrangement sequence of the edges is determined according to the out-degree and the in-degree of the connection nodes, and the storage format adopts a COO format, so that the storage efficiency of the nodes is improved to a greater extent. The rearranged edge sampling subgraph has better compactness in structure, so that the subgraph construction is more efficient, the GPU memory space utilization rate is improved, the time of each sampling iteration is reduced, and the GNN training time is accelerated.

The application discloses a graph processing system cache method, a controller and a system for a BFS algorithm, belongs to the technical field of performance optimization of a graph processing system, and comprises the following steps: based on the relationship between the access heat of a cache page and the number of unvisited vertices in the BFS algorithm, more accurate heat grading management is implemented on the cache page, when the cache is full, the page with low access potential is effectively eliminated, and the cache hit rate is improved; fragmented effective edge data is extracted from the cache page and separately stored in a cache managed in a fine granularity; by acquiring the proportion of the edge list of the vertex in the BFS algorithm task queue that has been resident in the cache, and in combination with the vertex out-degree and other attributes, the scheduling order of the vertex in the task queue is adjusted, and the data that has been resident in the cache is maximally utilized. The application can effectively improve the cache hit rate of the graph processing system when processing the BFS algorithm task, and further realizes the improvement of the access performance of the graph processing system.