[Spread-users] spread performance looks not very good for me

[Yair Amir]

Hi Ryan,

Your numbers also don't make much sense really.

1. There should be no decrease between the case of 1 sender and one (another)
receiver and the case of 1 sender and two (other, each on its own computer)
receivers if the network is one segment that is set up correctly.

2. 3 senders should be able to pump more than 1 sender. Did your 3
senders also receiving ? In that case this does make sense as each
sender is also receiving and in your first two tests it seems that
the sender is not receiving (am I correct?).

Also: I am not sure that the coarse-grain flow control in spflooder
allows running at full speed. In addition, I am not sure the default
flow control in Spread is set up for speed benchmarks. The default
parameters are set to optimize on the number of e-mails we get (as low
as possible) and not for best performance. Some of these parameters
can be twicked with spmonitor and some (esp. timeouts) require
recompiling.

Cheers,

       :) Yair.

On Friday, December 05, 2003 9:06 PM
Ryan Caudy caudy at jhu.edu wrote:

Ryan> I don't have a decent Solaris cluster to test on, but a quick test using 
Ryan> our Linux cluster (on a 100 Mbps ethernet network) gives (pretty rough) 
Ryan> results of:

Ryan> 1 sender + 1 receiver:  85 Mbps
Ryan> 1 sender + 2 receivers: 40 Mbps
Ryan> 3 senders:              52 Mbps

Ryan> In each of these tests, nodes listed as senders are receiving messages 
Ryan> as well, and nodes listed as receivers are just receiving.  Having more 
Ryan> sending nodes will increase the overall throughput, because more of the 
Ryan> token's circulation time will be used to send.  I'm not sure of the 
Ryan> details of the test that led to the 8000 * 1KB number on the Spread 
Ryan> Overview.

Ryan> I think the problem may lie outside of Spread, or with the Solaris port.

Ryan> Cheers,
Ryan> Ryan

Ryan> wei_hu at agilent.com wrote:

>> Spread performance question:
>> 
>> I am doing some investigation on spread to decide if it could be used as middleware in our project.  I studied spread for two days and did some benchmark.  I do have some questions.  Appreciate
>> any comment on them.
>> 
>> To do the benchmark, I used one Solaris machine as Sender and other one/two machines as Receivers.  All machines are in one subnet. Spfloor in spread package was used as benchmark program.  There
>> is average message rate I got.  Message size was 1K.
>> 1 sender + 1 receiver: 19.07 Mbps
>> 1 sender + 2 receiver: 13.46 Mbps
>> The sender and one of the receivers is 500Mhz CPU, 1G memory, Solaris OS.  Another receiver is 500Mhz CPU, 512M memory, Solaris OS.
>> 
>> My question is:
>> 1.    The message transmitting rate is far below what is claimed in "Spread Overview" (8K 1Kbytes per second).  Any possible reason for that?
>> 2.    Why plugging in more receivers decreased the message transmitting rate?
>> 3.    Spread daemon burned up to 70% CPU time?
>> 
>> _______________________________________________
>> Spread-users mailing list
>> Spread-users at lists.spread.org
>> http://lists.spread.org/mailman/listinfo/spread-users
>>