DMRlib: Easy-Coding and Efficient Resource Management for Job Malleability

Abstract

Process malleability has proved to have a highly positive impact on the resource utilization and global productivity in data centers compared with the conventional static resource allocation policy. However, the non-negligible additional development effort this solution imposes has constrained its adoption by the scientific programming community. In this work, we present DMRlib, a library designed to offer the global advantages of process malleability while providing a minimalist MPI-like syntax. The library includes a series of predefined communication patterns that greatly ease the development of malleable applications. In addition, we deploy several scenarios to demonstrate the positive impact of process malleability featuring different scalability patterns. Concretely, we study two job submission modes (rigid and moldable) in order to identify the best-case scenarios for malleability using metrics such as resource allocation rate, completed jobs per second, and energy consumption. The experiments prove that our elastic approach may improve global throughput by a factor higher than 3x compared to the traditional workloads of non-malleable jobs.

An Emulation Layer for Dynamic Resources with MPI Sessions

“…Malleability in MPI: Dynamic resource management in MPI has been studied intensively over the last years, usually under the umbrella term of "malleability". Since then, multiple frameworks have been created to support malleability in MPI applications [4,14,11,6,8]. These frameworks used different techniques and APIs to achieve malleability.…”

“…These frameworks used different techniques and APIs to achieve malleability. For example, some authors propose process splitting and merging for expansion and shrinking of the application [4,8]. Other authors start new processes while keeping the old, existing processes running [14].…”

Dynamic spawning of MPI processes applied to malleability

“…To date, malleability is infrequently used by applications, as described in Hori et al (2021) and Bernholdt et al (2018). Although checkpoint/restart (C/R) techniques exhibit similar behaviour and can ease the adoption of malleability in codes, they have exceedingly high reconfiguration times (Iserte et al, 2016).…”

“…The application of malleability has shown an approximately 20% reduction of the makespan in Posner and Fohry (2021). Furthermore, Iserte et al (2020) demonstrated the ways to use it to reduce makespan by approximately 4x when combined with malleability techniques, whereas Iserte and Rojek (2019) showed its impact on energy efficiency improving approximately 2.4x in GPU-capable workloads.…”

“…In Iserte et al (2020), the Dynamic Management of Resources (DMR) framework is introduced, which implements malleability using the OmpSs programming model 1 on the runtime Nanos++ 2 . This is a distributed parallel runtime based on MPI, responsible for handling processes and their communication.…”