Another way to scale-out server virtually: Build a big "logical" x86 server out of multiple relatively inexpensive x86 servers, instead of getting big Sun (ahem ... Oracle) or IBM SMP servers. Servers are connected via Infiniband. Unlike Vmware/Xen hypervisors, this method allows a workload to transparently span the resources of multiple physical servers. But even with Infiniband, the memory latency can still be very significant. I remember experiencing very poor scale-out performance on DB2 with the IBM x3950 multi-chassis setup.
Some good articles that describes virtual SMP solutions:
- http://ideasint.blogs.com/ideasinsights/2009/11/building-large-smp-systems-with-virtualization.html (
also mentioned a similar virtual SMP solution from 3Leaf that uses a proprietary server interconnect to reduce latenciesUPDATE - 3Leaf Systems is out of business) - http://virtualizationreview.com/blogs/everyday-virtualization/2010/04/smp-hypervisor-scalemp.aspx
- http://searchdatacenter.techtarget.com/news/2240024681/Move-over-SMPs-Distributed-shared-memory-systems-step-in
I would love to see some real world benchmarks of these virtual SMP solutions vs hardware optimized SMP servers.