Dear Tench:
For most applications that "Servers" do, a NUMA based computer can do as well if, not better. Thus a Beowulf Cluster of N machines can outrun a N way SMP box. That is because, 4 Xeons have to share each bus and the associated memory, while each Tbird has its own dedicated memory. Beowulf clusters have important failover capabilities that are not duplicated with SMP servers. You can lose a computer here and a computer there in a cluster and not stop processing user requests. In a SMP server, the whole thing shuts down when say a memory DIMM fails. This allows the cluster to use a rolling update, preventative maintenance, or backup. Other than to make scheduling easy, one can use hetrogeneous computers (read different speeds and sizes) in a cluster. In SMP, you can not. This allows a smooth upgrade (scaling) path, instead of SMP upgrades in large jumps. Also, the limit of CPUs is much larger in a cluster than SMP. 144 way clusters are common whereas 32 way servers are rare. Granted, SMP technology has more history, but clustering is highly sought after. DEC was the first to really push clustering. This technology was very profitable for DEC. Compaq now has this technology but, clustering is now in the public domain.
When you can get 72 1G Tbirds plus Beowulf Cluster infrastructure for less than a 8 way Xeon SMP server goes for, there will be much less demand for those servers.
Pete |
