• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • 2
  • Tagged with
  • 2
  • 2
  • 2
  • 1
  • 1
  • 1
  • 1
  • About
  • The Global ETD Search service is a free service for researchers to find electronic theses and dissertations. This service is provided by the Networked Digital Library of Theses and Dissertations.
    Our metadata is collected from universities around the world. If you manage a university/consortium/country archive and want to be added, details can be found on the NDLTD website.
1

Adaptive memory hierarchies for next generation tiled microarchitectures

Herrero Abellanas, Enric 05 July 2011 (has links)
Les últimes dècades el rendiment dels processadors i de les memòries ha millorat a diferent ritme, limitant el rendiment dels processadors i creant el conegut memory gap. Sol·lucionar aquesta diferència de rendiment és un camp d'investigació d'actualitat i que requereix de noves sol·lucions. Una sol·lució a aquest problema són les memòries “cache”, que permeten reduïr l'impacte d'unes latències de memòria creixents i que conformen la jerarquia de memòria. La majoria de d'organitzacions de les “caches” estan dissenyades per a uniprocessadors o multiprcessadors tradicionals. Avui en dia, però, el creixent nombre de transistors disponible per xip ha permès l'aparició de xips multiprocessador (CMPs). Aquests xips tenen diferents propietats i limitacions i per tant requereixen de jerarquies de memòria específiques per tal de gestionar eficientment els recursos disponibles. En aquesta tesi ens hem centrat en millorar el rendiment i la eficiència energètica de la jerarquia de memòria per CMPs, des de les “caches” fins als controladors de memòria. A la primera part d'aquesta tesi, s'han estudiat organitzacions tradicionals per les “caches” com les privades o compartides i s'ha pogut constatar que, tot i que funcionen bé per a algunes aplicacions, un sistema que s'ajustés dinàmicament seria més eficient. Tècniques com el Cooperative Caching (CC) combinen els avantatges de les dues tècniques però requereixen un mecanisme centralitzat de coherència que té un consum energètic molt elevat. És per això que en aquesta tesi es proposa el Distributed Cooperative Caching (DCC), un mecanisme que proporciona coherència en CMPs i aplica el concepte del cooperative caching de forma distribuïda. Mitjançant l'ús de directoris distribuïts s'obté una sol·lució més escalable i que, a més, disposa d'un mecanisme de marcatge més flexible i eficient energèticament. A la segona part, es demostra que les aplicacions fan diferents usos de la “cache” i que si es realitza una distribució de recursos eficient es poden aprofitar els que estan infrautilitzats. Es proposa l'Elastic Cooperative Caching (ElasticCC), una organització capaç de redistribuïr la memòria “cache” dinàmicament segons els requeriments de cada aplicació. Una de les contribucions més importants d'aquesta tècnica és que la reconfiguració es decideix completament a través del maquinari i que tots els mecanismes utilitzats es basen en estructures distribuïdes, permetent una millor escalabilitat. ElasticCC no només és capaç de reparticionar les “caches” segons els requeriments de cada aplicació, sinó que, a més a més, és capaç d'adaptar-se a les diferents fases d'execució de cada una d'elles. La nostra avaluació també demostra que la reconfiguració dinàmica de l'ElasticCC és tant eficient que gairebé proporciona la mateixa taxa de fallades que una configuració amb el doble de memòria.Finalment, la tesi es centra en l'estudi del comportament de les memòries DRAM i els seus controladors en els CMPs. Es demostra que, tot i que els controladors tradicionals funcionen eficientment per uniprocessadors, en CMPs els diferents patrons d'accés obliguen a repensar com estan dissenyats aquests sistemes. S'han presentat múltiples sol·lucions per CMPs però totes elles es veuen limitades per un compromís entre el rendiment global i l'equitat en l'assignació de recursos. En aquesta tesi es proposen els Thread Row Buffers (TRBs), una zona d'emmagatenament extra a les memòries DRAM que permetria guardar files de dades específiques per a cada aplicació. Aquest mecanisme permet proporcionar un accés equitatiu a la memòria sense perjudicar el seu rendiment global. En resum, en aquesta tesi es presenten noves organitzacions per la jerarquia de memòria dels CMPs centrades en la escalabilitat i adaptativitat als requeriments de les aplicacions. Els resultats presentats demostren que les tècniques proposades proporcionen un millor rendiment i eficiència energètica que les millors tècniques existents fins a l'actualitat. / Processor performance and memory performance have improved at different rates during the last decades, limiting processor performance and creating the well known "memory gap". Solving this performance difference is an important research field and new solutions must be proposed in order to have better processors in the future. Several solutions exist, such as caches, that reduce the impact of longer memory accesses and conform the system memory hierarchy. However, most of the existing memory hierarchy organizations were designed for single processors or traditional multiprocessors. Nowadays, the increasing number of available transistors has allowed the apparition of chip multiprocessors, which have different constraints and require new ad-hoc memory systems able to efficiently manage memory resources. Therefore, in this thesis we have focused on improving the performance and energy efficiency of the memory hierarchy of chip multiprocessors, ranging from caches to DRAM memories. In the first part of this thesis we have studied traditional cache organizations such as shared or private caches and we have seen that they behave well only for some applications and that an adaptive system would be desirable. State-of-the-art techniques such as Cooperative Caching (CC) take advantage of the benefits of both worlds. This technique, however, requires the usage of a centralized coherence structure and has a high energy consumption. Therefore we propose the Distributed Cooperative Caching (DCC), a mechanism to provide coherence to chip multiprocessors and apply the concept of cooperative caching in a distributed way. Through the usage of distributed directories we obtain a more scalable solution and, in addition, has a more flexible and energy-efficient tag allocation method. We also show that applications make different uses of cache and that an efficient allocation can take advantage of unused resources. We propose Elastic Cooperative Caching (ElasticCC), an adaptive cache organization able to redistribute cache resources dynamically depending on application requirements. One of the most important contributions of this technique is that adaptivity is fully managed by hardware and that all repartitioning mechanisms are based on distributed structures, allowing a better scalability. ElasticCC not only is able to repartition cache sizes to application requirements, but also is able to dynamically adapt to the different execution phases of each thread. Our experimental evaluation also has shown that the cache partitioning provided by ElasticCC is efficient and is almost able to match the off-chip miss rate of a configuration that doubles the cache space. Finally, we focus in the behavior of DRAM memories and memory controllers in chip multiprocessors. Although traditional memory schedulers work well for uniprocessors, we show that new access patterns advocate for a redesign of some parts of DRAM memories. Several organizations exist for multiprocessor DRAM schedulers, however, all of them must trade-off between memory throughput and fairness. We propose Thread Row Buffers, an extended storage area in DRAM memories able to store a data row for each thread. This mechanism enables a fair memory access scheduling without hurting memory throughput. Overall, in this thesis we present new organizations for the memory hierarchy of chip multiprocessors which focus on the scalability and of the proposed structures and adaptivity to application behavior. Results show that the presented techniques provide a better performance and energy-efficiency than existing state-of-the-art solutions.
2

Programming, debugging, profiling and optimizing transactional memory programs

Hasanov Zyulkyarov, Ferard 19 July 2011 (has links)
Transactional memory (TM) is a new optimistic synchronization technique which has the potential of making shared memory parallel programming easier compared to locks without giving up from the performance. This thesis explores four aspects in the research of transactional memory. First, it studies how programming with TM compares to locks. During the course of work, it develops the first real transactional application ¿ AtomicQuake. AtomicQuake is adapted from the parallel version of the Quake game server by replacing all lock-based synchronization with atomic blocks. Findings suggest that programming with TM is indeed easier than locks. However the performance of current software TM systems falls behind the efficiently implemented lock-based versions of the same program. Also, the same findings report that the proposed language level extensions are not sufficient for developing robust production level software and that the existing development tools such as compilers, debuggers, and profilers lack support for developing transactional application. Second, this thesis introduces new set of debugging principles and abstractions. These new debugging principles and abstractions enable debugging synchronization errors which manifest at coarse atomic block level, wrong code inside atomic blocks, and also performance errors related to the implementation of the atomic block. The new debugging principles distinguish between debugging at the language level constructs such as atomic blocks and debugging the atomic blocks based on how they are implemented whether TM or lock inference. These ideas are demonstrated by implementing a debugger extension for WinDbg and the ahead-of-time C# to X86 Bartok-STM compiler. Third, this thesis investigates the type of performance bottlenecks in TM applications and introduces new profiling techniques to find and understand these bottlenecks. The new profiling techniques provide in-depth and comprehensive information about the wasted work caused by aborting transactions. The individual profiling abstractions can be grouped in three groups: (i) techniques to identify multiple conflicts from a single program run, (ii) techniques to describe the data structures involved in conflicts by using a symbolic path through the heap, rather than a machine address, and (iii) visualization techniques to summarize which transactions conflict most. The ideas were demonstrated by building a lightweight profiling framework for Bartok-STM and an offline tool which process and display the profiling data. Forth, this thesis explores and introduces new TM specific optimizations which target the wasted work due to aborting transactions. Using the results obtained with the profiling tool it analyzes and optimizes several applications from the STAMP benchmark suite. The profiling techniques effectively revealed TM-specific bottlenecks such as false conflicts and contentions accesses to data structures. The discovered bottlenecks were subsequently eliminated with using the new optimization techniques. Among the optimization highlights are the transaction checkpoints which reduced the wasted work in Intruder with 40%, decomposing objects to eliminate false conflicts in Bayes, early release in Labyrinth which decreased wasted work from 98% to 1%, using less contentions data structures such as chained hashtable in Intruder and Genome which have higher degree of parallelism.

Page generated in 0.1056 seconds