Sun, 03 Sep 2006
We had problems with some users, who occasionally ran huge computing tasks on a general-purpose staff server, slowing down work for all other people (including mail server, home directories file serving, etc.). Because of this, and partly to test the technology as well, we have decided to order a separate server, which will be dedicated for such huge computing tasks (as well as for our experiments with big NUMA iron).
Yesterday the server has finally arrived. It is Tyan Transport VX50, an eight-CPU server with dualcore CPUs, 16 cores total. Apart from other things, we want to test Oracle on it. When we have bought the present IS MU database server, the biggest Opteron-based configurations had four CPUs. So we went for SGI Altix box (with 2 to 16 Itanium2 CPUs) instead. If the VX50 is stable and fast enough, our next DB server can probably be Opteron-based - who knows whether there will be SGI as a server manufacturer next year?
So far our VX50 has survived my stress-tests, including many parallel
kernel compiles, user-space variant of
load on ext3 and XFS, etc. The hardware works, including SATA controller,
hardware sensors, etc.
in a single-image system is pretty impressive, isn't it?
Probably the most interesting part of this server design is its NUMA topology (see above). They have managed to use all of the three HyperTransport channels of AMD 8xx CPUs, even though the configuration is not strictly symmetric then. I wonder how the routing of requests over the HT mesh is done. Does it use static routing, or some kind of load balancing? Any pointers, my dear lazyweb?