Post-graduate theses
Search command : Author="Αντωνίου"
And Author="Μαρία"
Current Record: 4794 of 6663
|
Identifier |
uch.csd.msc//2006kotsis |
Title |
Βελτιώνοντας την Κλιμακοσιμότητα σε Συστοιχίες Υπολογιστών Διαμοιρασμένης Μνήμης |
Alternative Title |
Improving Scalability on Shared Memory Clusters |
Creator |
Kotsis, George
|
Abstract |
Clusters of symmetric multiprocessors (SMPs) have become the predominant computing platform for high-performance parallel systems. Such clusters have traditionally used message passing communication. It is however widely argued that a system with shared memory is much easier to program than a system with message passing. For this reason, previous work has examined how a shared memory abstraction can be provided on top of clusters that do not offer native hardware support for shared memory. Current research initiatives examine new approaches to building cluster interconnects and communication protocols. It is not fully known how well, and at which performance level, these interconnects will be able to support a shared memory abstraction. In this work we examine the performance and scalability of an existing shared memory system (GeNIMA) on top of a cluster that uses edge-based interconnects. We port GeNIMA to the Hermes communication system that provides high-level communication semantics on top of an Ethernet-based interconnect. Hermes offers reliable and ordered remote read and write memory operations. Hermes is also able to transparently use multiple links for spatial parallelism and provides a novel API for allowing mixed in-order and out-of-order message delivery. We evaluate the behavior of GeNIMA and the underlying communication subsystem using eight real applications from the SPLASH-2 benchmark suite on a 32-nodes cluster interconnected with Ethernet topology. Our results show that: (i) Applications that perform reasonably well under shared memory can achieve speedups in the range of 8-24 on a 32 nodes configuration. (ii) The execution time of applications with high communication overheads improves up to 20% with two 1 Gigabit Ethernet links per node. (iii) Using a single 10 Gigabit Ethernet link results up to 20% improvement in the performance of communication demanding applications. (iv) In all configurations we examine, Hermes imposes very little extra traffic due to flow control actions.
|
Issue date |
2006-12-01 |
Date available |
2006-12-08 |
Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses
|
|
Type of Work--Post-graduate theses
|
Permanent Link |
https://elocus.lib.uoc.gr//dlib/e/4/b/metadata-dlib-2006kotsis.tkl
|
Views |
535 |