Research projects with DAS at Vrije Universiteit Amsterdam
This Java implementation runs extremely efficiently on DAS:
In the Albatross project we use DAS to
study algorithms and applications for wide-area parallel computing.
We have run a set of Orca applications on the wide-area DAS system,
analyzed their performance, and hand-optimized the applications to
take the wide-area structure into account. The results indicate
that surprisingly high efficiencies can be obtained on a wide-area cluster,
despite the high latencies and low throughput of the WAN links.
This project also focuses on wide-area parallelism; it uses MPI
for its communication layer. The new idea is to optimize MPI collective
operations, so they become aware of the wide-area architecture.
This way, MPI applications that use collective operations become
transparently optimized for wide-area usage.
We have developed efficient communication software for Myrinet (the local area
network of DAS), including LANai firmware, higher-level communication
protocols (Panda), and programming systems (Orca and PVM).
We have analyzed the performance of the Orca system on DAS in detail
and we have compared it against two other distributed shared memory
systems (CRL and TreadMarks).
We have also done a performance comparison
between Myrinet, ATM, and Fast Ethernet, using a collection
of challenging applications.
LFC is short for "Link-level Flow Control". It implements a flexible
low-level messaging layer, comparable in functionality to Active Messages,
Fast Messages or U-net.
- Native compiler
- Panda to
access the network (Myrinet, Fast Ethernet)
Back to the DAS home page
This page is maintained by Henri Bal.