Process Migration in FROST
Authors
Glibstrup, Michael ; Kringelbach, Lars
Term
10. term
Publication year
2002
Abstract
FROST er en distribueret, heterogen beregningsplatform, der lader forskellige typer computere dele og løse opgaver via et API. I dette projekt forbedrer vi FROSTs dynamiske belastningsfordeling ved at indføre procesmigrering, altså at flytte en kørende proces fra en hårdt belastet node til en mindre belastet under selve kørselen. Vi har analyseret, hvordan procesmigrering bedst kan udføres, og valgt en løsning baseret på checkpoints, som gemmer en proces’ tilstand, så den kan genoptages på en anden node. Løsningen er designet, og dele af designet er implementeret, med særlig fokus på ikke at forringe ydeevnen. Vi har testet systemet for overhead, samlet ydeevne og adfærden af migreringspolitikker (reglerne for, hvornår en proces flyttes). Testene viser, at den ekstra overhead ved procesmigrering er lille i forhold til gevinsterne, og at systemet fortsat performer godt med funktionen. For yderligere forbedringer anbefaler vi at overveje migreringspunkter, dvs. udtrykkelige steder hvor en proces sikkert kan flyttes.
FROST is a distributed, heterogeneous computing platform that lets different kinds of computers share and complete tasks via an API. In this project, we improve FROST’s dynamic load balancing by adding process migration, meaning a running process can be moved from a heavily loaded node to a less loaded one during execution. We analyzed how to perform process migration and chose a checkpoint-based approach that saves a process’s state so it can resume on another node. The solution was designed and selected parts were implemented, with careful attention to avoiding performance regressions. We evaluated overhead, overall performance, and the behavior of migration policies (the rules that decide when to move a process). Tests show the added overhead from process migration is small relative to its benefits, and the system continues to perform well with the feature. To further improve performance, we recommend considering migration points—explicit places where a process can be safely moved.
[This abstract was generated with the help of AI]
Documents