OpenDDS performance

[Malcolm Spence]

Hi everyone I wanted to post an update because I know many CORBA and
OMG communities
are also following OpenDDS as it leverages many of the same
technologies.

We have been getting the next version of OpenDDS ready for release it
has a
lot of improvements and more compliance. It will be out in few days.

We have also been doing some comprehensive testing of the new version,
replicating what we believe are essentially the circumstances that RTI
has
on their web site. (We could not find other DDS implementations' test
data.) Our test descriptions are here.

http://www.opendds.org/performance.html

Ironically we have a similar set of machines and network to RTI’s, in
our
training lab. 2.5 GHz systems on a 1 gig network. Our performance
numbers
overlay those of RTI. (We show more transports and so have a broad
spread.)

For a 1500 byte message size (which is a standard size for financial
market
data) we had 110 microseconds latency for “send to end” or one way,
from one
system to another across Ethernet.
http://www.opendds.org/perf/lab100125/latency_results.html

We then did some testing of OpenDDS on a single box using the SGI high
performance (3GHz systems) racked systems lab. In that we used the
single
loopback test approach. That gave us a baseline “cost” of OpenDDS and
that
was only 16 – 20 microseconds for the same 1500 byte message. (Of
course
shared memory transport would improve even that number, but that is
not the
point of this particular baseline testing.)

http://www.opendds.org/perf/sgi100127/latency_results.html

In other words the performance of any DDS is a function of language (C+
+
being faster than Java) then network topology, protocol and devices.
Proprietary technology does not buy you any improved performance once
you
have set up your physical circumstances.

We will be doing some additional testing across boxes in a “racked”
setting
using TCP/IP over Infiniband. We shall then see what additional
latency
accrues in a limited distributed systems arrangement. We expect it to
be
negligible. Even if it is noticeable, (Infiniband only adds only a few
hundred nanoseconds in its point to point switching), even that can be
mitigated by adding native Infiniband support as a transport protocol
(in
place of TCP/IP) in the OpenDDS transport framework. Then you have
basically a gigantic multiplexing capability with hundreds of systems
possible and short-message frame-times down around 20 microseconds.

I think OpenDDS is going to be an attractive option on many projects.

Regards Malcolm

Malcolm D. Spence
Director of Business Development
OCI "Use our reach to exceed your grasp."
Sun Authorized Java Center (tm)
Voting member of the OMG.
(Phone 1- 314-579-0066 ext. 206 or FAX -0065)
www.theaceorb.com www.ociweb.com