There have been some questions about how the worker machines really function. I have read what others have written but I still did not fully grasp it so I decided to try it out for myself using only IP addresses. I did figure out a few interesting things.
- If you only have WD users, adding only one worker machine is not really useful at all if it is comparable the the master in CPU/RAM. WD connections will be on the worker only unless the conditions in #2 and #3 are met.
- The master will only handle overflow WD connections if you are using hostnames/SSL.
- The master will only handle overflow WD connections if #2 is true and all worker machine have reached 100 connection limit.
I first did a test with one worker machine. Then I added a second worker machine. The master will not support any Web Direct clients if you have any worker machines deployed. So adding one worker machine when you only have WD clients is kind of the same as if you used the worker as the master. This is considering many deployments will have an average of 10-20 WD users. If you have 30+WD users I would guess that the master could be a modest 4 core machine and the worker could be a 20 core machine and there is a different thing going on there.
The master seemed to balance the load pretty evenly between the two worker machines. Adding multiple workers seems to be where the real benefit is even if you have only 10 WD users. When you hit the 100 connection limit on the workers FMS stops taking new WD connections and returns "A server with the specified hostname could not be found."
Per LisaRose, rollover to the master only happens when hostnames and SSL are used AND the worker machines all reach 100 connections. This might also be a way to reserve the master cpu for FMP and FMGo connections.
The redirect the master provides is seamless to the user. The redirection used about 5-7% of the master CPU resources on a 4 core server maintaining a user connection activity once per second.
Tested with the FMServer_Sample file.
Message was edited by: bigtom. Changed the original post based on info from @lisarose regarding physical test results using hostnames/SSL.