20072007 International Conference on Convergence Information Technology (ICCIT 2007)Requires access

Design and Implementation of an RDMA Gateway for Heterogeneous Clusters

Shin Gyu Kim, Hyuck Han, Hyungsoo Jung, Heon Young Yeom

Open publisher page 1 citations

Abstract

Building high-performance clusters using one of the two leading network technologies, Myrinet and InfiniBand, has been thought as a de facto way to achieve several teraflops computing power. Meanwhile, maintaining both types of clusters, it appears, may have created an another challenge for the MPI programming system, the most popular parallel programming library that has been successfully used on both networks. The belief that extending cluster resources across two different types of networks may increase computing parallelism has driven many researchers to tackle this challenge with various viewpoints. We approach this challenge with a different perspective, application transparency, which is accomplishing the goal without any modification of legacy MPI applications. We, therefore, focus on the design of an RDMA gateway that can relay messages very fast, and this design focus turns out to be a better way to preserve the application transparency. RDMA gateway (RG), our prototyped system, has a very efficient memory management mechanism that prevents RG from showing irregular spikes of a memory usage under a heavy load condition. Experimental results show that running parallel applications over heterogeneous clusters can be very promising with low performance overhead.

About this research paper

What this paper is about

Building high-performance clusters using one of the two leading network technologies, Myrinet and InfiniBand, has been thought as a de facto way to achieve several teraflops computing power. Meanwhile, maintaining both types of clusters, it appears, may have created an another challenge for the MPI programming system, the most popular parallel programming library that has been successfully used on both networks. The belief that extending cluster resources across two different types of networks may increase computing parallelism has driven many researchers to tackle this challenge with various viewpoints. We approach this challenge with a different perspective, application transparency, which is accomplishing the goal without any modification of legacy MPI applications. We, therefore, focus on the design of an RDMA gateway that can relay messages very fast, and this design focus turns out to be a better way to preserve the application transparency. RDMA gateway (RG), our prototyped system, has a very efficient memory management mechanism that prevents RG from showing irregular spikes of a memory usage under a heavy load condition. Experimental results show that running parallel applications over heterogeneous clusters can be very promising with low performance overhead.

Why it matters

OpenAlex reports 1 citations for this work. Citation counts describe recorded attention and do not establish research quality.

Key contribution

A contribution statement is not available in the OpenAlex record.

Method / approach

Method details are not available in the OpenAlex metadata.

Main findings

Findings are not separately available in the OpenAlex metadata.

Limitations

Limitations are not available in the OpenAlex metadata.

Applications

Application details are not available in the OpenAlex metadata.

Available abstract

Building high-performance clusters using one of the two leading network technologies, Myrinet and InfiniBand, has been thought as a de facto way to achieve several teraflops computing power. Meanwhile, maintaining both types of clusters, it appears, may have created an another challenge for the MPI programming system, the most popular parallel programming library that has been successfully used on both networks. The belief that extending cluster resources across two different types of networks may increase computing parallelism has driven many researchers to tackle this challenge with various viewpoints. We approach this challenge with a different perspective, application transparency, which is accomplishing the goal without any modification of legacy MPI applications. We, therefore, focus on the design of an RDMA gateway that can relay messages very fast, and this design focus turns out to be a better way to preserve the application transparency. RDMA gateway (RG), our prototyped system, has a very efficient memory management mechanism that prevents RG from showing irregular spikes of a memory usage under a heavy load condition. Experimental results show that running parallel applications over heterogeneous clusters can be very promising with low performance overhead.

Key concepts: Remote direct memory access, Myrinet, InfiniBand, Computer science, Transparency (behavior), Distributed computing, Low latency (capital markets), Default gateway

Related papers

Back to paper searchBrowse research topicsOriginal source
Design and Implementation of an RDMA Gateway for Heterogeneous Clusters — Research Paper | ScholarLens