English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Tools annual

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

Pipeline parallelism has been demonstrated to be a remarkable approach toimprove throughput for training deep neural networks with billions ofparameters over heterogeneous clusters. The 1F1B scheduling plan is a widelyadopted strategy for memory and performance optimization, which interchangesthe forward and backward stage computations of different micro-batches. On theother hand, a common issue in using the 1F1B scheduling is that stagecomputation is delayed due to the data transfer when network resources arepreempted by other tasks, even with the minimum communication between stages.The exclusive access of these network resources cannot be guaranteed in cloudofferings. We present a general scheduling technique to accommodate pipelineparallelism to preempted network environments at the expense of a certainamount of memory pressure. The core concept is to extend 1F1B schedule schemeto kFkB, which groups k micro-batches, and alternately executes k forward andbackward computations. We propose Ada-Grouper, an adaptive kFkB scheduler whichregularly adjusts the number of group members k to maintain an optimal balancebetween communication and computation efficiency correspond to changes in achanging network environment under the memory limit. Experimental resultsdemonstrate that our design maintain stable performance for pipelineparallelism, yielding a performance increase of up from 4% to 30%, comparedwith 1F1B in preempted network scenarios.

Ada-Grouper: Accelerating Pipeline Parallelism in Preempted Network by Adaptive Group-Scheduling for Micro-Batches