ADAPTIVE TASK ALLOCATION WITH DATA LOCALITY AWARENESS IN HETEROGENEOUS CLOUD NODES

Abstract

Data locality alone is insufficient for scheduling analytics tasks when cloud workers differ in computational capability and experience changing service rates. This paper presents Adaptive Locality aware Task Allocation (ALTA), an online dispatcher for independent, nonpreemptive tasks with replicated input blocks. The dispatcher combines a queue estimate, task dependent worker speed and input transfer time. Compute and transfer multipliers are updated from completed tasks using exponentially weighted moving averages. A bounded preference for local data selects the smallest remote input volume among placements whose predicted response time is within a specified additive budget of the minimum. The resulting policy requires a linear scan of feasible workers and admits a conditional one step error bound. A discrete event simulator evaluates seven policies under five synthetic operating conditions using 30 paired workload seeds and 1,200 tasks per run. Under combined compute and network drift, ALTA reduces mean response time from 6.60 to 5.10 s relative to static earliest finish time dispatch, a reduction of 22.8% with a paired bootstrap 95% interval of 20.4–25.3%. The mean per run 95th percentile decreases by 38.0%. Relative to adaptive earliest finish time dispatch without locality slack, remote input volume falls by 4.8%, while the response time difference is inconclusive. Stationary and network only experiments show no clear response time advantage. These results support adaptive service estimation and controlled locality preference within the stated simulation model; they do not establish superiority on production cloud workloads.