On Fri, Mar 08, 2024 at 06:06:45PM +0000, James Morse wrote:
Hi guys,
On 07/03/2024 23:16, Tony Luck wrote:
On Thu, Mar 07, 2024 at 02:39:08PM -0800, Reinette Chatre wrote:
Thank you for the example. I find that significantly easier to understand than a single number in a generic "nodes_per_l3_cache". Especially with potential confusion surrounding inconsistent "nodes" between allocation and monitoring.
How about domain_cpu_list and domain_cpu_map ?
Like this (my test system doesn't have SNC, so all domains are the same):
$ cd /sys/fs/resctrl/info/ $ grep . */domain* L3/domain_cpu_list:0: 0-35,72-107 L3/domain_cpu_list:1: 36-71,108-143 L3/domain_cpu_map:0: 0000,00000fff,ffffff00,0000000f,ffffffff L3/domain_cpu_map:1: ffff,fffff000,000000ff,fffffff0,00000000 L3_MON/domain_cpu_list:0: 0-35,72-107 L3_MON/domain_cpu_list:1: 36-71,108-143 L3_MON/domain_cpu_map:0: 0000,00000fff,ffffff00,0000000f,ffffffff L3_MON/domain_cpu_map:1: ffff,fffff000,000000ff,fffffff0,00000000 MB/domain_cpu_list:0: 0-35,72-107 MB/domain_cpu_list:1: 36-71,108-143 MB/domain_cpu_map:0: 0000,00000fff,ffffff00,0000000f,ffffffff MB/domain_cpu_map:1: ffff,fffff000,000000ff,fffffff0,00000000
This duplicates the information in /sys/devices/system/cpu/cpuX/cache/indexY ... is this really because that information is, er, wrong on SNC systems. Is it possible to fix that?
On an SNC system the resctrl domain for L3_MON becomes the SNC node instead of the L3 cache instance. With 2, 3, or 4 SNC nodes per L3.
Even without the SNC issue this duplication may be a useful convienience. On Intel to get from a resctrl domain is a multi-step process to first find which of the indexY directories has level=3 and then look for the "id" that matches the domain.
From Tony's earlier description of how SNC changes things, the MB controls remain
per-socket. To me it feels less invasive to fix the definition of L3 on these platforms to describe how it behaves (assuming that is possible), and define a new 'MB' that is NUMA scoped. This direction of redefining L3 means /sys/fs/resctrl and /sys/devices have different views of 'the' cache hierarchy.
I almost went partly in that direction when I started this epic voyage. The "almost" part was to change the names of the monitoring directories under mon_data from (legacy non-SNC system):
$ ls -l mon_data total 0 dr-xr-xr-x. 2 root root 0 Mar 8 10:31 mon_L3_00 dr-xr-xr-x. 2 root root 0 Mar 8 10:31 mon_L3_01
to (2 socket, SNC=2 system):
$ ls -l mon_data total 0 dr-xr-xr-x. 2 root root 0 Mar 8 10:31 mon_NODE_00 dr-xr-xr-x. 2 root root 0 Mar 8 10:31 mon_NODE_01 dr-xr-xr-x. 2 root root 0 Mar 8 10:31 mon_NODE_02 dr-xr-xr-x. 2 root root 0 Mar 8 10:31 mon_NODE_03
While that is in some ways a more accurate view, it breaks a lot of legacy monitoring applications that expect the "L3" names.
(I also think that this be over the threshold on 'funny machines look funny' - but I bet someone builds an arm machine that looks like this too!)
-Tony