[Spread-users] Trying to use spread on 6 machines in a lab.

Jonathan Stanton jonathan at cnds.jhu.edu
Mon Jun 24 23:17:51 EDT 2002

With the current release version the maximum number of daemons that will
work in one configuration (specified in a spread.conf file) is 128. This is
not just a compile constant (i.e. the protocol will require changes to
support more). So to support a large configuration there are two choices:

1) Run multiple configurations, each of which has a distinct subset of
nodes who do not need to communicate (or at least do not need globally
ordered messages) from any nodes not in their subset. We know people who
have done this successfully. You can have some hosts be part of multiple
configurations and if needed a client app can bridge some messages between
different configurations (with some loss in the strict semantics -- we
could talk about the details)

2) Run no more then 128 daemons, spread around as best as possible. Then
have most of the other clients connect to a deamon on a remote host. Since
each daemon can easily support almost 1000 clients (if they aren't too busy
each) a few daemosn can support a lot of clients, even on different
machines. Unless LAN bandwidth is a major bottleneck in your system, I
would recommend running no more then 30 daemons or so and having the rest
of the clients connect remotely. This is almost the same as your suggestion

If your WAN has large latency, there will be a noticable hit on the
performance of Spread (i.e. latency for each message will be higher and
flow control will need some tuning to optimize throughput).

If you have some specfic situation in mind, we could talk offline about
what options might be best.


On Mon, Jun 24, 2002 at 10:11:55AM -0400, Daniel J Walsh wrote:
> Is there any recommended way to use spread if you have a large 
> environment?  Say you have 1000 machines in 10 subnets over a Wide Area 
> Network.  Would you recommend a configuration with say two or more 
> spread servers (Machines running the spread daemon) in each subnet and 
> the clients built to fail over between the spread servers?
> Dan

Jonathan R. Stanton         jonathan at cs.jhu.edu
Dept. of Computer Science   
Johns Hopkins University    

More information about the Spread-users mailing list