Migration granularity

Given the existance of stuff such as RDMA and DSM (the
latter compliments of the MigSHM code), migration
granularity should be as fine or as coarse as you want
it or need it.

There has been mention at times of migrating
individual threads across the cluster (which can be a
valuable thing to do) but the technology to support
something at this level has massively improved since
it was first mooted. What is unclear is what work has
been done either to provide granularity control or to
use these newer mechanisms to support such control.

Does anyone know what the state of thread migration
is?

The other problem with migration is in deciding where
to migrate the process (or thread) to. As I understand
it, at present the decision system picks the node with
the greatest free time, which is valid only where all
nodes are of equal topological distance.

It is also incorrect to use the unweighted latency
and/or bandwidth between nodes, as not all processes
will carry out an equal amount of communication.

In the general case, the relative importance of CPU
cycles, network latency and network bandwidth will be
very application-specific.

It would therefore seem to me that the most correct
system is to default to the existing mechanism but to
allow a user to profile an application and have the
migration mechanism -optionally- use a more
comprehensive decision engine that can use those
profiles to plug in weights for the three parameters
and therefore migrate processes/threads to whatever
node closest meets the profiled requirements.

Again, any thoughts on how else we could tune the
cluster?


__________________________________________________
Do You Yahoo!?
Tired of spam?  Yahoo! Mail has the best spam protection around 
http://mail.yahoo.com 

-------------------------------------------------------------------------
Using Tomcat but need to do more? Need to support web services, security?
Get stuff done quickly with pre-integrated technology to make your job easier
Download IBM WebSphere Application Server v.1.0.1 based on Apache Geronimo
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642