On Sat, Oct 02, 2010 at 09:29:38AM +1000, Shane wrote: > I made the following comment in a thread on IBM-MAIN - perhaps the > Boeblingen folks might know. > <quote> > Given that Linux is (NUMA) node aware, it would be interesting to see > how a (non-z/VM) multi-book s390x Linux partition would appear re node > numbers and scheduler queues. My expectation would be that each book > would appear as a node - with its attendant memory. > I have no idea how z/VM may mangle this representation for a similarly > defined guest. > </quote>
The System z machines are no "real" NUMA machines since memory striping is used. Which means in a contiguous memory area each x MB will be from a different book. The NUMA effects will be partly hidden by the huge L3 (z10) / L4 (z196) caches. Or in other words: Linux is not aware if memory is local or remote. However the cpu topology is known (that is: which cpu belongs to which book) and the Linux kernel makes use of this in building scheduling domains (multi-core scheduling) accordingly and then tries to keep processes on a book instead of migrating them back and forth. With z196 another cache level was introduced. Linux kernel support for this should be merged upstream with the upcoming 2.6.37 kernel. ---------------------------------------------------------------------- For LINUX-390 subscribe / signoff / archive access instructions, send email to lists...@vm.marist.edu with the message: INFO LINUX-390 or visit http://www.marist.edu/htbin/wlvindex?LINUX-390 ---------------------------------------------------------------------- For more information on Linux on System z, visit http://wiki.linuxvm.org/