Look into Mosix. It's a kernel patch that load balances processes across the cluster, much like in an SMP. Keep in mind, though, that all network and file IO (IIRC) stays local to the login node that the process is started on, so you'll definitely want to have a very fast switched network, and a fast disk subsystem (RAID) on the login node.
I'm not sure how this would really help a single HLDS though. If a single HLDS process eats all of a single CPU in 1 cluster node, and all the CPUs in the cluster are the same speed, then Mosix will probably not migrate the process to another node. What mosix will do, however, is load balance all the HLDS processes across the available nodes automatically, thus alleviating the need to manually assign (start) HLDS locally on each node. If you were to try this, and since the login node typically will have 2 NICs (1 public, 1 private cluster only)), you'll need to start each HLDS server on a different port, as all the public traffic will transit a single NIC (single IP unless you alias). They may have some new IO distibution in the latest versions, but I haven't kept up with it. Keep in mind, this is all only (somewhat) educated conjecture on my part, as I've not tried it. Read all about Mosix here: http://www.mosix.org StanTheMan TheHardwareFreak http://www.hardwarefreak.com rcon admin at: Beer for Breakfast servers <http://bfb.bogleg.org/> 209.41.98.2:27016 (CS multi-map) 209.41.98.2:27015 (DoD) 209.41.98.2:27017 (CS militia/dust2) Dallas, TX > -----Original Message----- > From: David Touitou [mailto:[EMAIL PROTECTED]] > Sent: Sunday, September 01, 2002 6:52 AM > To: [EMAIL PROTECTED] > Subject: RE: [hlds_linux] Anyone Ever Try & Use Beowulf To Run a HLDS > Serv er > > > Quoting [EMAIL PROTECTED]: > > > Short answer: > > An nice definition, thanks Stan. > > > Think of a cluster system, including Beowulf, Mosix, MPI, PVM, etc, > > as an SMP with slower interprocess communication. > > Clusters work well on parallel codes, but not on serial codes, > > unless you have a lot of serial codes you need to run at the same > > time. > > That's my thought (and Beowulf or PVM won't help, we need another type > of cluster here) : considering a cluster as a "giant" SMP (one or two > processors per node) and having several HLDS processes > jumping from one > CPU to the CPU (ie from one node to the other) depending on > the load of > each node. > That's the way HLDS behave on a standard SMP system. > > But you need a powerfull "scheduler" (correct me if the name > is wrong) to getthe HLDS processes jump from nodes to nodes. > And high bandwidth on the cluster front-end. > > > Thus, there might be an administrative advantage running HLDS on a > > cluster, if and only if, you run many HLDS servers. > > I'm talking hundreds. > > Not that sure you need hundreds to have a gain. You'll get > advantage as > soon as you're in the trouble I've talked about yesterday. > > David. > > -- > http://gamez.solexine.fr/ > > _______________________________________________ > To unsubscribe, edit your list preferences, or view the list > archives, please visit: > http://list.valvesoftware.com/mailman/listinfo/hlds_linux > _______________________________________________ To unsubscribe, edit your list preferences, or view the list archives, please visit: http://list.valvesoftware.com/mailman/listinfo/hlds_linux
