Knowledge graph analytics kernels in high performance computing
Data mining large-scale corpora of scholarly publications, such as the full biomedical literature, which may consist of tens of millions of papers spanning decades of research. The present disclosure provides a Distributed Accelerated Semiring All-Pairs Shortest Path (DSNAPSHOT) algorithm for computing shortest paths of a knowledge graph using distributed-memory parallel computers accelerated by GPUs. DSNAPSHOT implementations can analyze connected input graphs with millions of vertices using a large number graphics processing units (e.g., the 24,576 GPUs of the Oak Ridge National Laboratory's Summit supercomputer system). DSNAPSHOT provides sustained performance of about 136*10 15 floating-point operations per second (136 petaflop/s) at a parallel efficiency of about 90% under weak scaling and, in absolute speed, 70% of the performance given our computation (in the single-precision tropical semiring or “min-plus” algebra). DSNAPSHOT may enable mining of scholarly knowledge corpora when embedded and integrated into artificial intelligence-driven natural language processing workflows at scale.
1. A method, performed by a high-performance computing system, to enable mining of existing information and deriving new knowledge, the method comprising:
by the high-performance computer system, receiving, from multiple textual databases information about a set of articles and annotations relating to biomedical concepts that appear in the articles;
by the high-performance computing system, forming a first knowledge graph using the entire received information and annotations, the knowledge graph having nodes connected by edges, where the nodes include
a set of article nodes representing respective biomedical articles, and
a set of concept nodes representing respective biomedical concepts;
wherein edges between the concept nodes represent relations between biomedical concepts based on co-occurrence of biomedical terms in the articles,
wherein edges between the concept nodes and the article nodes represent relations between biomedical concepts and biomedical articles based on annotations or mentions within the biomedical articles, and
wherein edges between the article nodes represent relations between biomedical articles based on citation references;
by a set of processors of the high-performance computing system, determining shortest paths between all pairs of nodes of the first knowledge graph, the computing system performing the shortest-path computations by:
implementing a distributed Floyd-Warshall (FW) algorithm across multiple processing units in a two-dimensional process grid;
coordinating parallel updates using a message passing interface (MPI) that performs diagonal updates, panel updates, and MinPlus Outer Product computations over a tropical semiring executed by graphics processing units (GPUs);
performing computations in CPU-only mode, GPU-only mode, or CPU-GPU overlapping mode to optimize resource utilization based on workload;
performing asynchronous lookahead scheduling to overlap computation and communication; and
broadcasting matrix updates using a bandwidth-optimal ring communication protocol;
placing communicating processes based on optimal rank placement and performing intra-node communication optimization to reduce communication latency;
by the high-performance computing system, forming a second knowledge graph that includes the nodes of the first knowledge graph and the edges of the first knowledge graph that correspond only to the shortest paths;
by the high-performance computing system, further optimizing communication load by placing communicating processes on processing units that are physically proximal within the high-performance computing system; and
by the high-performance computing system, predicting yet-undiscovered biomedical relationships between the biomedical concepts based on the second knowledge graph, wherein the predictions are computed using the shortest paths as input features; and
outputting a ranked list of biomedical relationships based on the shortest-path computations to a display.
2. The method of claim 1 , further comprises applying, while determining the shortest paths, a constraint to at least one of downweight, remove at least some edges along particular paths, and ignore a shortest path completely.
3. The method of claim 2 , wherein the constraint represents a measure of relevance between concepts, between articles, or between concepts and articles.
4. The method of claim 1 , wherein communications for determining the shortest paths is performed using only CPU communicators, only GPU communicators, or concurrently both CPU-GPU communicators of the high-performance computing system.
5. The method of claim 1 , wherein the high-performance computing system uses GPU communicators when the distributed FW algorithm is performed only by GPUs.
6. The method of claim 5 , further comprising reducing communication burden by instructing communicating subprocesses to run on resources of the high-performance computing system that are closer together.