> (AGP/PCI-E, PCI-X, all onboard sound/video/sata/pata/network, > etc) connected thru CPU0. The only thing CPU1 (zero based, so the second > CPU) has hanging off it is the memory belonging to its memory controller. > That's the way my Tyan s2885 is laid out. > > > I run irqbalance, and irq's are evenly spread on both CPU's. Same > > temperature difference > > Of course, if I did something like that on mine, it would cause potential
This is an excellent point. IRQ balance may show that the IRQs are handled evenly across the CPUs, but as much of the motherboard I/.O traffic has to go through the CPU0 interconnect, that may heat it up. Another compicating factor: Numa. I have Numa enabled in my kernel, but others may not. This will affect the memory access traffic, that may be directed to a specific processor depending upon load. For example without Numa enabled, most memory accesses may come, for small memory loads, from the lower half of memory, which may be attached to only one of the processors. But if "chip kill" is enabled, these accesses are striped across the CPUs. Thus the accesses are spread. Has anyone actually put a calibrated thermometer against a reference portion of the CPU? I haven't... I am not convined that "sensors.conf" does anything more that try to indicate what the i2c channels are for certainly I haven't seen any incantations that might imply calibration..... Moreover gkrellm is only partially effective at picking up these indications, I often have to "tweak" the settings, and refer directly to the motherboard manual and plug and unplug connectors! Jason. -- gentoo-amd64@gentoo.org mailing list