[Sis-ams] teleconferencing the AMS meetings

[Scott Burleigh]

As promised, here is a summary of the *extremely preliminary* AMS 
performance benchmarking I've done to date.  These are results I got 
from using AMS to publish 200 million bytes of data in various 
configurations (lots of tiny messages, a smaller number of larger 
messages, etc.), with either 1 subscriber or 2 subscribers.  The numbers 
in the "2 subscr" column are per subscriber, so the total performance is 
twice the number shown.  The publishing application was running on a 
Pentium III-based machine running Red Hat 8.0 Linux; the subscribing 
application(s) ran on a second Pentium III-based machine that was 
running Fedora Core 3.  The two machines were privately connected via a 
100-Mbps Ethernet switch with no other traffic; AMS was running over a 
TCP transport system.

* *

	

* *

	

*Msgs/sec*

	

*Msgs/sec*

	

*Mbps*

	

*Mbps*

*Msg Size*

	

*Msg Count*

	

*1 subscr*

	

*2 subscr*

	

*1 subscr*

	

*2 subscr*

20

	

10 million

	

11560

	

8581

	

1.7

	

1.3

200

	

1 million

	

10124

	

6642

	

15.4

	

10.1

2000

	

100,000

	

4800

	

2529

	

73.2

	

38.6

20000

	

10,000

	

587

	

293

	

89.6

	

44.8

Notice how the performance goes from being (apparently) CPU-limited for 
the very small messages to I/O-limited for the largest messages.  The 
peak message rate I saw was about 17,000 total messages per second, so 
it was taking on average a little less than 60 microseconds to issue 
each message.

Plenty of room for optimization (e.g., the code wasn't even compiled 
with any optimization switches set), performance will be different on 
different hardware, and it will be interesting to see how a 
message-queue-based system does in contrast to TCP.

Scott

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.ccsds.org/pipermail/sis-ams/attachments/20050910/4bdad4fb/attachment.htm