On Thu, 29 Jan 2004 16:19:57 +0000 "Eugene de Villiers" wrote: [cc'ing back to suse-amd. Please keep discussions there]

...

Do a head -30 /var/log/boot.msg

dm(229) head -30 /var/log/boot.msg Inspecting /boot/System.map-2.4.21-193-smp Loaded 14927 symbols from /boot/System.map-2.4.21-193-smp. Symbols match kernel version 2.4.21. Loaded 117 symbols from 5 modules. klogd 1.4.1, log source = ksyslog started. ok <4>Bootdata ok (command line is root=/dev/sda2 vga=0x31a desktop splash=silent) <4>Linux version 2.4.21-193-smp (root@x86_64.suse.de) (gcc version 3.3.1 (SuSE Linux)) #1 SMP Wed Jan 21 18:41:44 UTC 2004 <6>BIOS-provided physical RAM map: <4> BIOS-e820: 0000000000000000 - 000000000009fc00 (usable) <4> BIOS-e820: 000000000009fc00 - 00000000000a0000 (reserved) <4> BIOS-e820: 00000000000e0000 - 0000000000100000 (reserved) <4> BIOS-e820: 0000000000100000 - 000000009fff0000 (usable) <4> BIOS-e820: 000000009fff0000 - 000000009ffff000 (ACPI data) <4> BIOS-e820: 000000009ffff000 - 00000000a0000000 (ACPI NVS) <4> BIOS-e820: 00000000ff7c0000 - 0000000100000000 (reserved) <4>kernel direct mapping tables upto 10100000000 @ 8000-d000 <6>Scanning NUMA topology in Northbridge 24 <3>Node 0 using interleaving mode 1/0 <6>No NUMA configuration found <6>Faking a node at 0000000000000000-000000009fff0000 <4>Bootmem setup node 0 0000000000000000-000000009fff0000 <7>ACPI: have wakeup address 0x10000002000 <4>Scan SMP from 0000010000000000 for 1024 bytes. <4>Scan SMP from 000001000009fc00 for 1024 bytes. <4>Scan SMP from 00000100000f0000 for 65536 bytes. <6>found SMP MP-table at 00000000000ff780 <4>hm, page 000ff000 reserved twice. <4>hm, page 00100000 reserved twice. <4>hm, page 000f9000 reserved twice.

Hmm, very strange, no sizes reported. Why is it using SMP and not NUMA for memory sharing? Is there a NUMA version of the 193 kernel I could use?

The kernel is NUMA. It cannot access the northbridge of the second CPU, so NUMA support is turned off. That's some hardware problem at your side, maybe the second CPU or the motherboard are broken or not correctly plugged in. -Andi

On Thu, 29 Jan 2004 20:12:54 +0100 Frank Pieczynski wrote:

There is a BIOS setting called "BIOS -> AML ACPI table", which is set to "Enabled" Helptext: "Set this value to allow the ACPI BIOS to add a pointer to an OEMB table in the Root System Description Table (RSDT). Note: OEMB table is used to pass POST data to the AML code during ACPI O/S operations."

That is not related. The Linux/x86-64 NUMA code doesn't use ACPI at all.

In memory config "Bank interleaving" and "Node interleaving" are disabled.

Bank Interleaving can be enabled, just Node interleaving should be disabled. The northbridge configuration says node interleaving is enabled though. Or maybe you only have DIMMs in the slots of one CPU. It's possible that the BIOS then sets the interleaving mode, although it's pointless. It wouldn't make a difference in that case. Anyways, the performance difference is not that big so I wouldn't worry too much about it.

May be this explains, why xengine displays only 200rpms (on a Matrox G550).

No, it shouldn't impact 3d performance much.

Some problem in ACPI implementation or BIOS bug?

Nothing to do with ACPI. The memory controller is configured by the BIOS. -Andi