On Mon, May 31, 2021 at 10:36:01AM +0200, Michal Privoznik wrote:
We allow configuring HMAT for domains since v6.6.0-rc1~249 (and friends). Basically, HMAT is more fine grained description of interconnects of NUMA nodes than basic NUMA distances. The former describes bandwidths and latencies while the latter is some dimensionless and normalized number. Anyway, mgmt apps did not really know what values to set for HMAT because we are not exposing them in capabilities because we were waiting on kernel to expose them. And it just did. In 09/10 I'm describing sysfs interface briefly and also mention that there's no interpretation of links to memory side caches, yet. I'm talking to kernel developers so we might get some movement there. But also, I'm not sure whether it's worth the effort OR if there really is a machine that has separate links to main memory and caches. Here's link to ACPI spec: https://uefi.org/sites/default/files/resources/ACPI_6_2.pdf Look for "5.2.27.4 System Locality Latency and Bandwidth Information Structure". And here's link to sysfs docs: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/Documentation/admin-guide/mm/numaperf.rst Michal Prívozník (10): tests: glib-ify vircaps2xmltest schemas: Allow zero <cpu/> for capabilities capabilities: Separate <cpu/> formatting into a function numa_conf: Rename virDomainCache* to virNumaCache* numa_conf: Expose virNumaCache formatter capabilities: Expose NUMA memory side cache numa_conf: Rename virDomainNumaInterconnect* to virNumaInterconnect* numa_conf: Expose virNumaInterconnect formatter
To these 8 patches above: Reviewed-by: Martin Kletzander <mkletzan@xxxxxxxxxx>
capabilities: Expose NUMA interconnects vircaps2xmltest: Introduce HMAT test case
These last two need some fix and/or explanation.
Attachment:
signature.asc
Description: PGP signature
