[Pmem] [PATCH 13/15] memory tiering: extend promotion threshold range

From: Huang Ying &lt;ying.huang(a)intel.com&gt; ANBZ: #80 cherry-picked from https://git.kernel.org/pub/scm/linux/kernel/git/vishal/tiering.git/commit/?… In current implementation, to adjust the memory tiering promotion threshold more quickly, we use the fixed steps (16) to adjust. But now we can reduce the threshold (more strictly) much more quickly, so we can use some other way to adjust the threshold to enlarge the range of the threshold and to make the threshold more accurate. In this patch, if the number of the pages that pass the threshold is more than 110% of the target number, we will decrease the threshold to 90% of its original value until 1ms. And we will increase the threshold to 110% of its original value vice versa. In this way, the minimal threshold is 1 ms (vs. 62ms by default originally). And the possible values of threshold becomes much more. Both make it possible to adjust the threshold more accurately. One downside is that if the PMEM pages become much colder suddenly, we will need more time to adjust the threshold to adapt to that. So that the promotion throughput will be lower than expected. But if the PMEM pages become colder, it’s less urgent to promote them too. So it should be OK to promote them slower. Eventually, the threshold can be adjusted to the proper value. In the test with the pmbench memory accessing benchmark on a 2-socket server machine with Optane DCPMM, the promotion threshold can be adjusted to a smaller value (33 ms vs. 62 ms originally). The pmbench score decreases 2.8%, mainly because the promotion throughput decreases 4.9%. Because the number of the pages that can be promoted isn't distributed evenly vs. time, but the ratelimit is enforced equally vs. time. Will resolve the issue with the following patches in the patchset. Signed-off-by: "Huang, Ying" &lt;ying.huang(a)intel.com&gt; Signed-off-by: Baolin Wang &lt;baolin.wang(a)linux.alibaba.com&gt; --- kernel/sched/fair.c | 16 ++++++++-------- 1 file changed, 8 insertions(+), 8 deletions(-) diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 5806316..74dec2e 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -2927,14 +2927,12 @@ int sysctl_numa_balancing_threshold(struct ctl_table *table, int write, void *bu return err; } -#define NUMA_MIGRATION_ADJUST_STEPS 16 - static void numa_migration_adjust_threshold(struct pglist_data *pgdat, unsigned long rate_limit, unsigned long ref_th) { unsigned long now = jiffies, last_th_ts, th_period; - unsigned long unit_th, th, oth; + unsigned long th, oth; unsigned long last_nr_cand, nr_cand, ref_cand, diff_cand; th_period = msecs_to_jiffies(sysctl_numa_balancing_scan_period_max); @@ -2952,13 +2950,15 @@ static void numa_migration_adjust_threshold(struct pglist_data *pgdat, } pgdat->numa_threshold_ts = now; pgdat->numa_threshold_nr_candidate = nr_cand; - unit_th = ref_th / NUMA_MIGRATION_ADJUST_STEPS; oth = pgdat->numa_threshold; th = oth ? : ref_th; - if (diff_cand > ref_cand * 11 / 10) - th = max(th - unit_th, unit_th); - else if (diff_cand < ref_cand * 9 / 10) - th = min(th + unit_th, ref_th); + if (diff_cand > ref_cand * 11 / 10) { + th = min(th * 9 / 10, th - 1); + th = max(th, 1UL); + } else if (diff_cand < ref_cand * 9 / 10) { + th = max(th * 11 / 10, th + 1); + th = min(th, ref_th * 2); + } pgdat->numa_threshold = th; spin_unlock(&pgdat->numa_lock); trace_autonuma_threshold(pgdat->node_id, diff_cand, th); -- 1.8.3.1