On Fri, 27 Oct 2023, Maciej Wieczór-Retman wrote:
> On 2023-10-24 at 12:26:26 +0300, Ilpo Järvinen wrote:
> >CAT test spawns two processes into two different control groups with
> >exclusive schemata. Both the processes alloc a buffer from memory
> >matching their allocated LLC block size and flush the entire buffer out
> >of caches. Since the processes are reading through the buffer only once
> >during the measurement and initially all the buffer was flushed, the
> >test isn't testing CAT.
> >
> >Rewrite the CAT test to allocate a buffer sized to half of LLC. Then
> >perform a sequence of tests with different LLC alloc sizes starting
> >from half of the CBM bits down to 1-bit CBM. Flush the buffer before
> >each test and read the buffer twice. Observe the LLC misses on the
> >second read through the buffer. As the allocated LLC block gets smaller
> >and smaller, the LLC misses will become larger and larger giving a
> >strong signal on CAT working properly.
> >
> >The new CAT test is using only a single process because it relies on
> >measured effect against another run of itself rather than another
> >process adding noise. The rest of the system is allocated the CBM bits
> >not used by the CAT test to keep the test isolated.
> >
> >Replace count_bits() with count_contiguous_bits() to get the first bit
> >position in order to be able to calculate masks based on it.
> >
> >This change has been tested with a number of systems from different
> >generations.
> >
> >Suggested-by: Reinette Chatre <reinette.chatre@xxxxxxxxx>
> >Signed-off-by: Ilpo Järvinen <ilpo.jarvinen@xxxxxxxxxxxxxxx>
> >---
> > tools/testing/selftests/resctrl/cat_test.c | 286 +++++++++-----------
> > tools/testing/selftests/resctrl/fill_buf.c | 6 +-
> > tools/testing/selftests/resctrl/resctrl.h | 5 +-
> > tools/testing/selftests/resctrl/resctrlfs.c | 44 +--
> > 4 files changed, 137 insertions(+), 204 deletions(-)
> >
> >diff --git a/tools/testing/selftests/resctrl/cat_test.c b/tools/testing/selftests/resctrl/cat_test.c
> >index e71690a9bbb3..7518c520c5cc 100644
> >--- a/tools/testing/selftests/resctrl/cat_test.c
> >+++ b/tools/testing/selftests/resctrl/cat_test.c
> >@@ -11,65 +11,68 @@
> > #include "resctrl.h"
> > #include <unistd.h>
> >
> >-#define RESULT_FILE_NAME1 "result_cat1"
> >-#define RESULT_FILE_NAME2 "result_cat2"
> >+#define RESULT_FILE_NAME "result_cat"
> > #define NUM_OF_RUNS 5
> >-#define MAX_DIFF_PERCENT 4
> >-#define MAX_DIFF 1000000
> >
> > /*
> >- * Change schemata. Write schemata to specified
> >- * con_mon grp, mon_grp in resctrl FS.
> >- * Run 5 times in order to get average values.
> >+ * Minimum difference in LLC misses between a test with n+1 bits CBM mask to
> >+ * the test with n bits. With e.g. 5 vs 4 bits in the CBM mask, the minimum
> >+ * difference must be at least MIN_DIFF_PERCENT_PER_BIT * (4 - 1) = 3 percent.
> >+ *
> >+ * The relationship between number of used CBM bits and difference in LLC
> >+ * misses is not expected to be linear. With a small number of bits, the
> >+ * margin is smaller than with larger number of bits. For selftest purposes,
> >+ * however, linear approach is enough because ultimately only pass/fail
> >+ * decision has to be made and distinction between strong and stronger
> >+ * signal is irrelevant.
> > */
> >-static int cat_setup(struct resctrl_val_param *p)
> >-{
> >- char schemata[64];
> >- int ret = 0;
> >-
> >- /* Run NUM_OF_RUNS times */
> >- if (p->num_of_runs >= NUM_OF_RUNS)
> >- return END_OF_TESTS;
> >-
> >- if (p->num_of_runs == 0) {
> >- sprintf(schemata, "%lx", p->mask);
> >- ret = write_schemata(p->ctrlgrp, schemata, p->cpu_no,
> >- p->resctrl_val);
> >- }
> >- p->num_of_runs++;
> >-
> >- return ret;
> >-}
> >+#define MIN_DIFF_PERCENT_PER_BIT 1
> >
> > static int show_results_info(__u64 sum_llc_val, int no_of_bits,
> >- unsigned long cache_span, unsigned long max_diff,
> >- unsigned long max_diff_percent, unsigned long num_of_runs,
> >- bool platform)
> >+ unsigned long cache_span, long min_diff_percent,
> >+ unsigned long num_of_runs, bool platform,
> >+ __s64 *prev_avg_llc_val)
> > {
> > __u64 avg_llc_val = 0;
> >- float diff_percent;
> >- int ret;
> >+ float avg_diff;
> >+ int ret = 0;
> >
> > avg_llc_val = sum_llc_val / num_of_runs;
> >- diff_percent = ((float)cache_span - avg_llc_val) / cache_span * 100;
> >+ if (*prev_avg_llc_val) {
> >+ float delta = (__s64)(avg_llc_val - *prev_avg_llc_val);
> >
> >- ret = platform && abs((int)diff_percent) > max_diff_percent;
> >+ avg_diff = delta / *prev_avg_llc_val;
> >+ ret = platform && (avg_diff * 100) < (float)min_diff_percent;
> >
> >- ksft_print_msg("%s Check cache miss rate within %lu%%\n",
> >- ret ? "Fail:" : "Pass:", max_diff_percent);
> >+ ksft_print_msg("%s Check cache miss rate changed more than %.1f%%\n",
> >+ ret ? "Fail:" : "Pass:", (float)min_diff_percent);
>
> Shouldn't "Fail" and "Pass" be flipped in the ternary operator? Or the condition
> sign above "<" should be ">"?
I must not touch ret ? "Fail:" : "Pass:" logic, it's the correct way
around. If I'd touch it, it'd break what the calling code assumes about
the return value.
(More explanation below).
> Now it looks like if (avg_diff * 100) is smaller than the min_diff_percent the
> test is supposed to fail but the text suggests it's the other way around.
>
> I also ran this selftest and that's the output:
>
> # Pass: Check cache miss rate changed more than 3.0%
> # Percent diff=45.8
> # Number of bits: 4
> # Average LLC val: 322489
> # Cache span (lines): 294912
> # Pass: Check cache miss rate changed more than 2.0%
> # Percent diff=38.0
> # Number of bits: 3
> # Average LLC val: 445005
> # Cache span (lines): 221184
> # Pass: Check cache miss rate changed more than 1.0%
> # Percent diff=27.2
> # Number of bits: 2
> # Average LLC val: 566145
> # Cache span (lines): 147456
> # Pass: Check cache miss rate changed more than 0.0%
> # Percent diff=18.3
> # Number of bits: 1
> # Average LLC val: 669657
> # Cache span (lines): 73728
> ok 1 CAT: test
>
> The diff percentages are much larger than the thresholds they're supposed to
> be within and the test is passed.
No, the whole test logic is changed dramatically by this patch and
failure logic is reverse now because of it. Note how I also altered these
things:
- MAX_DIFF_PERCENT -> MIN_DIFF_PERCENT_PER_BIT
- max_diff_percent -> min_diff_percent
- "cache miss rate within" -> "cache miss rate changed more than"
The new CAT test measures the # of cache misses (or in case of L2 CAT
test, LLC accesses which is used as a proxy for L2 misses). Then it takes
one bit away from the allocation mask and repeats the measurement.
If the # of LLC misses changes more than min_diff_precent when the
number of bits in the allocation was changed, it is a strong indicator CAT
is working like it should. Based on your numbers above, I'm extremely
confident CAT works as expected!
I know for a fact that when the selftest is bound to a wrong resource id
(which actually occurs on broadwell's with CoD enabled without one of the
later patches in this series), this test is guaranteed to fail 100%,
there's no noticeable difference measured in LLC misses in that case.
> >@@ -143,54 +168,64 @@ static int cat_test(struct resctrl_val_param *param, size_t span)
> > if (ret)
> > return ret;
> >
> >+ buf = alloc_buffer(span, 1);
> >+ if (buf == NULL)
>
> Similiar to patch 01/24, wouldn't this:
> if (!buf)
> be better?
I've already changed this based on the comment you made against 1/24 :-).
--
i.
