RE: [SSI-users] Questions about process migration
Brought to you by:
brucewalker,
rogertsang
From: Walker, B. J <bru...@hp...> - 2004-01-28 04:00:33
|
Josh, I'm not sure there are any specific migration numbers at this point. However, I can help to answer the questions. The current migration code does not start executing the process on the new node until all it's dirty data has been copied off the old node (note this is all dirty data, which might not be the same as the footprint). Assuming you had 100M of dirty data, it would take no less than 1 second to move on a 1gig and probably at least 2 seconds. If this is our of range for your app, you'll have to wait for someone to do the on-demand page movement (which would allow the process to start executing as soon as it gets the task structure over and then demands over pages as needed), which no one is working on at this point. 2. TCP sockets are automatically kept open whether they are on cvip or not. However, understand that the socket data structures stay on the node where they were created. Processes move and they can still access objects (sockets, pipes, etc.) that they created before they migrated, but the objects themselves don't move. 3. I'm not sure about mixed architecture issues. By default we like to run the same kernel on each node and that can handle some level of heterogeneity. It is possible to configure with different kernels on each node if that is necessary. However, there is no code to handle doing rexecs and migrates between different architectures so if that is necessary (different register sets, etc.), then we can't support it. Bruce > -----Original Message----- > From: ssi...@li...=20 > [mailto:ssi...@li...] On=20 > Behalf Of Josh > Sent: Tuesday, January 27, 2004 2:40 PM > To: ssi...@li... > Subject: [SSI-users] Questions about process migration >=20 >=20 > Hello all,=20 >=20 > I have high hopes of being able to deploy OpenSSI for my=20 > application in the next week or two. The processes I'm=20 > going to need migrated are very sensitive to latency and=20 > also need to keep their network connections (tcp sockets)=20 > intact after migration. So to make sure this project is=20 > viable, I would certainly appreciate answers to these=20 > questions (and any other advice of course :).=20 >=20 > 1)About how long should I expect process migration to take=20 > with a process with about 100M memory footprint. Assume=20 > a GigE backend and mid-end modern systems (ie low end dual=20 > Xeon, or decent HT P4).=20 >=20 > 2)Is it possible to keep the tcp socket open while using the=20 > virtual IP of the cluster? What's involved in setting that up?=20 >=20 > 3) A side question what are the issues involved with using=20 > a mixed architecture cluster? (ie P4 with AMD, or P4 with=20 > P3 etc).=20 >=20 > Thanks!=20 >=20 > -josh bardt >=20 >=20 > ------------------------------------------------------- > The SF.Net email is sponsored by EclipseCon 2004 > Premiere Conference on Open Tools Development and Integration > See the breadth of Eclipse activity. February 3-5 in Anaheim, CA. > http://www.eclipsecon.org/osdn > _______________________________________________ > Ssic-linux-users mailing list > Ssi...@li... > https://lists.sourceforge.net/lists/listinfo/ssic-linux-users >=20 |