Frog Asynchronous Graph Processing on GPU with Hybrid Coloring Model

By | June 21, 2018

Abstract:

GPUs have been progressively used to quicken chart preparing for muddled computational issues with respect to diagram hypothesis. Numerous parallel diagram calculations receive the offbeat processing model to quicken the iterative union. Shockingly, the predictable nonconcurrent figuring requires locking or nuclear activities, prompting huge punishments/overheads when actualized on GPUs. Accordingly, the shading calculation is embraced to isolate the vertices with potential refreshing clashes, ensuring the consistency/accuracy of the parallel preparing. Normal shading calculations, in any case, may experience the ill effects of low parallelism on account of an extensive number of hues by and large required for preparing an expansive scale diagram with billions of vertices.

Existing System:

We propose a light-weight nonconcurrent preparing system called Frog with a preprocessing/cross breed shading model. The crucial thought depends on the Pareto rule (or 80-20 control) about shading calculations as we saw through masses of certifiable chart shading cases. We find that a greater part of vertices (around 80 percent) are hued with just a couple of hues, to such an extent that they can be perused and refreshed in a high level of parallelism without abusing the consecutive consistency. As needs are, our answer isolates the handling of the vertices in view of the dispersion of hues. In this work, we for the most part answer three inquiries: (1) how to segment the vertices in a meager chart with expanded parallelism, (2) how to process substantial scale diagrams that can’t fit into GPU memory, and (3) how to diminish the overhead of information exchanges on PCIe while handling each parcel.

Proposing System:

We direct investigations on true information (Amazon, DBLP, YouTube, RoadNet-CA, WikiTalk, and Twitter) to assess our approach and make examinations with surely understood non-preprocessed, (for example, Totem, Medusa, MapGraph, and Gunrock) and preprocessed (Cusha) approach, by testing four traditional calculations (BFS, PageRank, SSSP, and CC). On all the tried applications and datasets, Frog can fundamentally beat existing GPU-based diagram preparing frameworks with the exception of Gunrock and MapGraph. MapGraph improves execution than Frog when running BFS on RoadNet-CA. The correlation amongst Gunrock and Frog is uncertain. The frog can beat Gunrock in excess of 1.04X when running PageRank and SSSP, while the benefit of Frog isn’t evident when running BFS and CC on some datasets particularly for RoadNet-CA.

Leave a Reply

Your email address will not be published. Required fields are marked *