E-Locus - Institutional Repository of the University of Crete - Analysis and optimizayion of overheads in communication protocols over high speed ethernet-based cluster interconnects

Home Analysis and optimizayion of overheads in communication protocols over high speed ethernet-based cluster interconnects

Results - Details

[Add to Basket]

Identifier

000323148

Title

Analysis and optimizayion of overheads in communication protocols over high speed ethernet-based cluster interconnects

Author

Πασσάς, Σταύρος Ιωάννη

Select a value

Μπίλας, Αντώνης

Abstract

At the core of contemporary high performance computer systems is the communication infrastructure. For this reason, there has been a lot of work on providing low-latency, high-bandwidth communication subsystems for clusters. In this work, we introduce MultiEdge, a connection oriented communication system designed for high-speed commodity hardware. We use MultiEdge to examine the behavior of edge-based protocols. We examine the implications of building a single logical link out of multiple physical links and we see how overheads and performance scale with the number and speed of links. Finally, we examine the costs associated with data copying at the 15-30 GBits/s range. We implement and show the effectiveness of zero-copy data transfers, analyzing the impact of additional costs associated with this approach.
We find that:
(a) Our base protocol reaches almost 99.2% of the nominal aggregate throughput for up-to 2 psychical links of 1 GBit/s link rate.
(b) When up-to 8 physical links are used, our protocol achieves up-to 65% of the nominal aggregate throughput.
(c) The impacts of interrupts and data copies are significant, and when they are removed, protocol reaches 80% and 100% of the nominal throughput respectively.
(d) With bi-directional link rates of 30 GBits/s, data copies limits the effective throughput to about 8.8 GBits/s on our systems.
(e) The use of our zero-copy mechanism results in 80-90% improvement and reaches up-to 57% of the available bi-directional throughput.
Finally, we believe that efficiently distributing protocol processing over multiple host CPUs emerges as the main challenge in achieving higher transfer rates on modern architectures.

Physical description

xi, 48 σ. : εικ. ; 30 cm.

Language

English

Issue date

2008-04-11

Collection

School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses

Type of Work--Post-graduate theses

Views

533

Digital Documents
	Download document View document Views : 4