Thread (331 messages) 331 messages, 12 authors, 2024-02-13

RE: [PATCH v5 8/8] selftests/resctrl: Adjust effective L3 cache size when SNC enabled

From: "Luck, Tony" <tony.luck@intel.com>
Date: 2023-09-19 14:36:32
Also in: linux-patches, lkml

On a system that has SNC disabled the function reports both "node_cpus"
and "cache_cpus" equal to 56. In this case snc_ways() returns "2". It is
the same on a system with SNC enabled that reports the previously mentioned
variables to be different by a factor of two (36 and 72).
Is it possible for node_cpus and cache_cpus to not be multiples of each
other? (as in for example cache_cpus being 10 and node_cpus being 21?).
If not I'd suggest using "==" instead of ">=".
Some CPUs may be offline when the test is run. E.g. with one CPU offline on SNC
node 0, you'd see node_cpus = 35 and cache_cpus = 71. But with one CPU offline
on node 1, you'd have node_cpus = 36, cache_cpus = 71.


If yes then I guess something like this could work? :
+     if (node_cpus >= cache_cpus)
+             return 1;
+     else if (2 * node_cpus >= cache_cpus)
+             return 2;
+     else if (4 * node_cpus >= cache_cpus)
+             return 4;

This returns "4" for the 36 71 case. But should still be "2".
quoted
PS. I did my tests on two Intel Ice Lakes.
Perhaps easier to play with the algorithm in user code?

#include <stdio.h>
#include <stdlib.h>

static int snc(int node_cpus, int cache_cpus)
{
     if (node_cpus >= cache_cpus)
             return 1;
     else if (2 * node_cpus >= cache_cpus)
             return 2;
     else if (4 * node_cpus >= cache_cpus)
             return 4;
     return -1;
}

int main(int argc, char **argv)
{
        printf("%d\n", snc(atoi(argv[1]), atoi(argv[2])));

        return 0;
}

N.B. it's probably not possible to handle the case where somebody took ALL the CPUs in SNC
node 1 offline (or SNC nodes 1,2,3 for the SNC 4 case).

I think it reasonable that the code handle some simple "small number of CPUs offline" cases.
But don't worry too much about cases where the user has done something extreme.

-Tony

Keyboard shortcuts
hback out one level
jnext message in thread
kprevious message in thread
ldrill in
Escclose help / fold thread tree
?toggle this help