Description
Date depot: 1 janvier 1900
Titre: Garbage collection for managed runtime on multicore processors
Directeur de thèse:
Marc SHAPIRO (LIP6)
Directeur de thèse:
Gaël THOMAS (SAMOVAR)
Encadrant :
Julien SOPENA (LIP6)
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Non defini
Resumé:
Managed Runtime Environments (MREs), such as a JVM or a CLI, are
increasingly successful for application development and execution,
mainly because they provide good properties of manageability, isolation,
and safety in general. A key component is the automated memory manager
or garbage collector (GC). Garbage collection automatically frees
unused memory, relieving application programmers from complex protocols
and guaranteeing the absence of memory leaks or access violations.
However, this comes at a price. Experimentally, GC may add up to 20% or
30% to mutator (i.e., application) execution time.
This performance penalty is particularly an issue with the recent
emergence of multicore architectures. Indeed, current GC algorithms
used in production MREs environments execute sequentially and block all
mutator threads. Amdhal's law predicts that, due to this sequential
bottleneck, even hundreds or thousands of cores will not speed up a
mutator any better than a factor of five. The performance of mutators
cannot scale with the number of cores.
Furthermore, the performance impact of GC on multicore architectures is
exacerbated by the memory architecture. Indeed, current GC algorithms
destroy cache locality, leading to numerous cache misses, and further
decreasing performance.
Doctorant.e: Gidra Lokesh