From: Nuno S. <nun...@vg...> - 2003-08-09 22:32:34
|
Hello! I know nothing about mosix but I'll answer anyway (so don't take my word on it). stevegt@TerraLuna.Org wrote: > Partially answering myself -- after building a new host kernel with the > migshm patch and deploying it on a cluster, I've so far been unable to > get any UML process to migrate, so one or more of the five "assumptions" > below apparently isn't being met. Can anyone spot which one(s)? > > Steve > AFAIK, UML don't use posix shared mem. You can make it use tmpfs - which is the base for linux's shmem - for it's "memory file", but that's it. Again, AFAIK uml processes won't migrate because they're better in their home node: - it's where the "disk" is (the image file); - it's where the "nic" is (the TUN device et al...); - it's where the "ram" is (the /tmp mmaped file); Also, UML requires lot's of host's ptrace activity... I don't think you can migrate ptraced processes easily... Can you try something for me? In your mosix setup run: strace seq 1 1000000000 > /dev/null This will be CPU bound... Then try to migrate the "seq" process. Does it work? Regards, Nuno Silva P.S. I know nothing about *mosix, so this questions may be totally dumb :) > On Fri, Aug 08, 2003 at 06:37:29PM -0700, stevegt@TerraLuna.Org wrote: > >>Hi All! >> >>Can anyone on the user-mode-linux-devel list verify whether UML guests >>fit the five assumptions (below) made by the openMosix shared memory >>migration patch? If so, this would be important -- UML guests could run >>on top of an openMosix cluster, with atuomatic migration of UML guest >>processes between nodes for load balancing. >> >>Details below. For more information, see http://mcaserta.com/maask/. >> >>Steve >> >> >>Release Notes - Migshm-1.4 >>------------- >> >>Migshm is an extension to openMosix which enables migration of >>shared memory processes. This patch will enable migration of processes >>using SYSV shared memory through shmget(), shmat(), shmdt(), shmctl() >>system calls. Also threads created using clone() system call can be >>migrated using migshm. >> >>The patch has been ported to the openMosix-2.4.21-1 kernel. >> >>Assumptions: >>----------- >> >>1. Shared memory applications use semaphores to acheive >>synchronization among the processes using shared memory. >> >>2. Applications that do not need synchronization, will use the flush() >>system call to acheive cluster-wide consistency. >> >>3. At any point of execution, processes are attached to only one >>shared memory region (This can be fixed and will be done soon). >> >>4. For thread migration, the parent remains on the home node while the children >>can migrate anywhere. >> >>5. The parent (and child) thread do not malloc() any memory, after the clone() >>call has been executed. The parent can malloc() before creating the child though. >> >>For both, migration of processes using SYSV shared memory and thread >>migration, the patch handles consistency of the shared data >>transparently. Also frequencies of access to the shared memory are >>being logged. Thus depending on the usage of shared memory and current >>load on the cluster, migshm enables migration of the shared memory >>itself and migration of group of processes. Migshm is still in its >>development phase and the known bug list is given below for reference: >> >>Known Bug List: >>-------------- >>(Work is going on. To be released in the next version) >> >>1. If a shared memory process migrates and does a remote shmdt(), then >>the refernce count of the shared memory is decremented on the remote >>but not on the home node. So on the home node the shared memory on the >>home node will not get destroyed. >> >>2. If you run the "ab" tester for Apache server, in a while(true) loop it >>breaks the home-node 2 times after a very long time. Also, no DEBUG messages >>printed under the DEBUG option on error conditions appear. These crashes >>may most likely have their reason in a general overload of the head-node/cluster. >>At least that is what I assume. >> --- As reported by Matthais Rechenburg >> >>Contact Information: >>------------------- >>Any feedback, suggestions, bug reports or test reports are >>welcome. You can mail us at maaskmaask @ hotmail.com. >> >> >> >>-- >>Stephen G. Traugott (KG6HDQ) >>UNIX/Linux Infrastructure Architect, TerraLuna Aerospace LLC >>stevegt@TerraLuna.Org >>http://www.stevegt.com -- http://Infrastructures.Org >> >> >>------------------------------------------------------- >>This SF.Net email sponsored by: Free pre-built ASP.NET sites including >>Data Reports, E-commerce, Portals, and Forums are available now. >>Download today and enter to win an XBOX or Visual Studio .NET. >>http://aspnet.click-url.com/go/psa00100003ave/direct;at.aspnet_072303_01/01 >>_______________________________________________ >>User-mode-linux-devel mailing list >>Use...@li... >>https://lists.sourceforge.net/lists/listinfo/user-mode-linux-devel > > |