Mike, This sounds like a perfect use for a hardware load-balancer. A hardware load-balancer can create the affinity for a particular server much like you describe - except for the choosing of a server based on your algorithm, that is.
Most hardware solutions implement some sort of persistence feature where, once a load-balancing decision has been made by the device, it will try to persist that same 'user' to that same server for a configurable amount of time. We use this to keep a user's access and application log on one server versus having to hunt around all the servers to piece their session together. If you are trying to create logic that says, "All project 'A' users go on server 'X'," I see that as a big challenge for the chicken-egg scenario you describe - the user already has an affinity to a server before you know the user or the project. However, there are some pretty slick hardware load-balancers out there that can use cookie-driven values. Check out the lineup from F5 - specifically their Big-IP Local Traffic Managers. Of course, you could have a farm of login servers that simply takes the login info and then redirects the user to a different farm of servers based on the project info. Good luck, and let us know what you come up with. + jay On 7/11/07, Mike Wynholds <[EMAIL PROTECTED]> wrote: > > > > > I have a somewhat complex load balancing scenario that I wish to accomplish. > From my investigation it seems that it's *probably* possible but I haven't > gotten it to work yet. Here is the scenario: > > > > I have a web app running on Resin 3.1.1 with Apache 2.2 in front of it. I > want to have multiple Resin servers in a cluster. I may have multiple > Apache servers as well, but I don't think that's important. If I do I will > have a hardware load-balancer in front of that. > > > > I want to achieve both failover and server scalability with my systems > architecture. I think I can utilize a clustering solution to handle this. > I will need to cluster both Resin sessions as well as a number of > memory/disk caches using the distributed mode of EhCache (or possible > JBossCache). So far the Resin session clustering is working fine. However, > for efficiency I really also want to segment my data between users, in order > to reduce either a) the amount of cache replication (if I use cache > replication) or b) the amount of cache misses (if I use cache invalidation) > in EhCache. > > > > When people come to the site, they are given a login page immediately. > There are two entry points: one is HTML-based and one is Flash/Flex-based. > Either way the user is prompted to log in with a username, password and a > project number. I want to pin all users logged in to a particular project > to the same backend server so that the EhCache can be used as efficiently as > possible. > > > > But while I say I want to "pin" specific users to specific backend servers, > I really also want the failover capability built in to load-balancers. So > what I *really* want is a way to suggest to the load balancer which server > to prefer, but allow it to fall back in the case of a server being down. My > algorithm for choosing a server to suggest is simple: I just hash the > project number, mod by the amount of servers, and choose that one. Not > smart, but for now that's fine. > > > > I feel like I will want to use the Resin load balancing features (although I > am open to other suggestions). From the docs it seems like there may be a > way to do it. I have spent a few hours on this and am not seeing a very > clear solution. My main theoretical issue is that when the user hits the > login page a session is created. There is no way to know which project > number the user will then enter, and therefore which server to try to pin > him too. But then if he enters a project number, Resin already has him > pinned to a server by the jssesionid. > > > > Is there a way to tell Resin to keep the session but now house it on a > different server? > > > > I think the logic of project number à server needs to be in the app > somewhere (ideally the server side, not the Flash side). Can I set a cookie > or something that can be sniffed by Resin's load balancing capabilities that > will determine the suggested server to pin to? > > > > I can kind of see the light at the end of the tunnel but haven't been able > to come up with even a best theoretical approach to this problem. There is > just too much about Resin load balancing, mod_caucho, hardware load > balancers and the such that I am not intimately familiar with, and I don't > want to reinvent anything that already exists and works well. > > > > Thanks for any suggestions. > > ..mike.. > > > > ..... > > Michael Wynholds > > President > > Carbon Five, Inc. > > 310 821 7125 x13 > > [EMAIL PROTECTED] > > > _______________________________________________ > resin-interest mailing list > resin-interest@caucho.com > http://maillist.caucho.com/mailman/listinfo/resin-interest > > _______________________________________________ resin-interest mailing list resin-interest@caucho.com http://maillist.caucho.com/mailman/listinfo/resin-interest
