> My findings are that persistent is offering great benefits, havnt tried an excessively harsh test yet, but i'm about to do that.
Just ran sniffer in both persistent and non-persistent modes with over 1,000 mesages in the overflow and MaxQueProc at 50. This pegs out my CPU between 90% & 100% for the duration of delivery.
Screenshots & sniffer log snipplets at http://staff.netsmith.net/sniffer/Extreme_Load/ I wont waste the mailing lists bandwith for the attachments for those who dont want them.
I dont see an obvious different when the system is under heavy load, at least not by skimming the log files.
Could do some math on overall performance statistics I guess... # of messages processed in same timeframe, average times, etc.
I see just a bit of difference in the margins - and in particular in recovery times. There are fewer long strings of large setup numbers in the persistent mode.
In theory, cellular peer-server based systems approach the performance of true client-server systems as system loads increase. The biggest differences show up on the low end. Once things get really busy there is no time for any server instances to rest so at least one will always be present.
Nailing up a persistent instance suppresses extra work on the low end and in the mid range so that a given system can handle more traffic before the load starts to effect performance - and also leaves plenty of resources open through the entire low end of the curve.
At the extreme high-end of the curve the persistent instance variant tends to perform better because the colony spends less effort on reorganizing itself. A purely peer-server implementation will begin to show signs of strain and edge toward catastrophic states earlier in high load scenarios as reorganization tasks begin to add random noise to the traffic flow and increase noise generated losses in the queue.
A persistent instance model suppresses these effects until the limits of performance are reached on the equipment. After that, both systems will fail in some spectacular ways - but usually this is masked by other problems before the effects can be demonstrated.
--- so goes the theory.
Anyway, if you squint just a little bit you can see a shade more daylight at the top of the cpu graph for the persistent test (IMO - yeah, I'm biased.)
Thanks again for this data! _M
PS: For grins I checked the latest rulebase compile record. During these tests each message was tested against 24656 heuristics.
This E-Mail came from the Message Sniffer mailing list. For information and (un)subscription instructions go to http://www.sortmonster.com/MessageSniffer/Help/Help.html
