 | PDA-Version |
 |
Home |
|
About Us |
|
Research |
|
Publications |
|
Contact |
|
Impressum |
| |
| Disclaimer/ Haftungsausschluss |
| Webmaster (admin@lfbs...) |
| |||||||
MPI for SCI-connected Clusters | |||||||
Documents
This is the first paper on SCI-MPICH. It is a complete description of the implementation and includes a theoretical performance model as well as the basic performance measurements.
The MP-MPICH Manual is contained in the "doc" directory of the release.
This talk deals, of course, with the same topic as the paper with the same name, but has some more actual additions. Some NAS benchmark results (also compared with ScaMPI) are included, and a prelimitary measurement of MPI-IO via SCI.
This paper documents the development SCI-MPICH has undergone since it was first published. Next to increased stability, usability and compatibility, the major new features are fully asynchronous message transfers via DMA and Interrupts and the introduction of MPI-IO via SCI. The slides of the talk are also available. [PDF] Efficient Asychronous Message Passing via SCI with Zero-Copying [PDF]
Zero-Copying via SCI is common for memory regions allocated via the SCI driver, but could not yet be realized for arbirary memory regions allocated by the user. This paper explains the implementation of generic zero-copying techniques for the SCI-MPICH, a freely available implementation of MPI for SCI connected clusters, and evaluates the performance benefits. slides of the talk [PDF] [PowerPoint] Exploiting Transparent Remote Memory Access for Non-Contiguous and One-Sided-Communication [PDF]
This paper presents two of the most recent optimizations in SCI-MPICH: efficient communication with non-contiguous MPI datatypes and one-sided communication according to the MPI-2 standard. We show that the transparent low-latency communication characteristics of SCI provides these techniques an excellent platform. slides of the talk [PDF] [PowerPoint] Student PapersA number of students have made contributions to SCI-MPICH. Regretfully, all of these paper are written in german language (because we are located in Germany).
(Optimization of the Transmission of Non-Contiguous Data in an MPI-Library for SCI-Clusters)) In this paper, a new technique, together with a smart and efficient algorithm, is presented which vastly improves the performance for the transmission of arbitrary, non-contiguous MPI datatypes on shared-memory and SCI-connected platforms. Such an optimization is not known for any other MPI implementation. It is also described in another paper (submitted for publication).
/Optimization and Evaluation of a Message Protocol for Memory-Coupled PC-Clusters) [ask me if you want to read it] This paper describes the optimization of the SCI-MPICH "short" protocol to make better use of the available memory resources while at the same time reducing the effective message latency. The memory blocks for the messsages do no longer have a fixed size, but are allocated as required.
(Definition, Implementation and Evaluation of an Asynchronous Communication Protocol for Memory-Coupled PC-Clusters using DMA and Interrupts) [ask me if you want to read it] The first version of the asynchronous, DMA-based rendez-vus protocol is described in this paper. In the mean time, this protcol has been extended towards a real zero-copy variant. | |||||||
| Author: NIL |
Print Version  |
| Related Links |
|---|
| MetaMPICH |
| NT-MPICH |
| SCI-MPICH |
| Performance |
| MP-MPICH |
| Mailinglist |
|---|
| We have installed an MP-MPICH mailing list for announcements and discussion. To subscribe to this list, send a mail to the mailinglist manager or use the online form. |