> I am involved in discussions with some UNIX guys that insist > Linux on s390 > is not "vertically scalable". I need to get around this > point. I have > some questions.
There is some truth to this assertion. Linux on S/390 does not do well where solving the problem is gated on delivering volumes of raw CPU cycles to a single instance of Linux -- the 390 CPUs are designed for balanced workloads with lots of changes in workload and priorities, not raw compute power on individual streams of instructions. Applications which expect this behavior do not do well on 390. I would stress with them that many, if not most, applications no longer rely on this behavior -- most have been split into multiple tiers (the WWW server/apps server/database model) and monolithic applications support is less important than balacing a large number of differing workloads in the operations area. That said, here's some of the not-so-true pieces: > I have heard that a single Linux image under VM can schedule > a maximum of 4 > CPU's without serious performance degradation. Is this true? With Linux 2.2, this is true, but it is a limitation of Linux, not the 390. The 2.4 kernels do pretty well with 8 CPUs, and are acceptable with up to about 16. Performance drops off beyond that point to a degree that makes it not very attractive. > Would this vertical scalability be offset in any way by the > capability of > running multiple images on they same physical processor? Yes. Much of the value is in operations cost savings. Applications that can really use a large n-way MP efficiently are fairly rare. There's a lot more money spent on maintaining multiple systems (in part because applications *don't* use the HW efficiently) than the value of vertical scalability. > Would this setup > need the applications cooperation for data sharing or not? The applications would need to share their data just as they do today if they are running on different systems. It's often more effective to have a number of small instances running with data sharing in place than one megasystem. If they happen to share the same hardware, that's a plus. > How do we compare processor power between HP, Sun and IBM > with relation to > UNIX and Linux? In general, you don't, because it's comparing apples to kumquats. Different tradeoffs were made in the designs of the processors, and comparing MIPS or "processor power" just gets you into a black maze of "what is comparable performance". The I/O system and operations cost are more interesting discussions, especially as we start seeing a lot more convergence in I/O. I favor looking at the outcome of the environment as a whole, including management and personnel costs as metrics, as a better comparison of how effective a solution is, but YMMV. Barton Robinson differs with me here, but I'll let him explain his view, as he does a better job of it. -- db
