6. oktober 2008 06:05I have 5 8-core servers and very coarse-grained, embarrassingly parallel, compute-intensive (minutes to days) problems to solve...and I'd like to use the fancy new HPC SOA model.
So, I'd like to use all 5 systems as Compute nodes, with the first also acting as a Head and Broker node (again, not a big deal theoretically, since my needs are very lightweight in traffic and management).
However, it seems that HPC Server Manager won't allow the head node to perform all three roles...does this mean, in order to accomplish my goal, that I'd be required to install the broker and/or the head in a virtual machine sitting on the first box? If so, what topology would you suggest (ie. 5 "pure" compute nodes and a virtual head/broker)? Is there a work-around with config files?
23. oktober 2008 20:15bump - anyone?
And a follow-up: If there's no work-around besides virtualization, would I be better off with a virtualized head or virutalized router node? How would I predict the relative workload of head and router nodes for compute-intensive or bandwidth-limited SOA jobs?
24. oktober 2008 01:43There is no work around for the constraint that a node can be either a compute node or broker node but not both.
As for trading of virtual vs physical, I would lean towards make the head node physical rather than virtual for the following reasons:
a) if you are running RRAS (nat) or DHCP (assigning addresses on a private network) we have found these services to be simpler to configure and more reliable on physical machines in a mixed mode cluster.
b) if the host for the virtual machine is a compute node and the vm gets starved by a compute intensive task then you may see communication failures talking to the vm or timeouts talking to sql - this could lead to the job failing.
c) if you use the vm as a compute node you can sandbox the resources that it consumes in order to ensure that the head node is still responsive.
d) if you are bandwidth limited for SOA then you are better off running it on the phsyical node (there is still some overhead for networking in the vm)
Developer, Microsoft HPC.
- Markeret som svar af Don PatteeModerator 22. maj 2009 20:53